An evaluation of the feasibility, cost and value of information of a multicentre randomised controlled trial of intravenous immunoglobulin for sepsis (severe sepsis and septic shock): incorporating a systematic review, meta-analysis and value of information analysis

Definitions of severe sepsis and septic shock

Sepsis is a syndrome characterised by a systemic inflammatory response to infection that leads to rapid acute organ failure and potentially rapid decline to death. Sepsis, severe sepsis and septic shock are generic terms and do not represent a single homogeneous disease; rather they are terms for a common syndrome.

In an attempt to formalise a definition for the sepsis syndrome, in 1991, a consensus conference was convened by the American College of Chest Physicians (ACCP) and the Society of Critical Care Medicine (SCCM). 1 This conference defined the concept of the systemic inflammatory response syndrome (SIRS) – a systemic activation of the innate immune response, regardless of cause. SIRS could be triggered by multiple insults, including infection, trauma, burns and pancreatitis. SIRS was considered to be present if two or more of the following four specific conditions were satisfied:

• temperature > 38°C or < 36°C

• heart rate > 90/min

• respiratory rate > 20/min or partial pressure of carbon dioxide (PaCO₂) < 4.3 kPa, and

• white blood cell count > 12 × 10⁹/l or < 4 × 10⁹/l (or > 10% immature neutrophils – ‘bands’).

Sepsis was defined as SIRS (above) in response to infection, severe sepsis as sepsis associated with organ dysfunction, hypoperfusion or hypotension and septic shock as sepsis with hypotension despite adequate fluid resuscitation (Figure 1). These definitions have formed the basis of entry criteria to the majority of recent studies investigating sepsis.

FIGURE 1.

Definitions of sepsis, severe sepsis and septic shock.

In 2001, another consensus conference was convened, sponsored by the SCCM, the European Society of Intensive Care Medicine (ESICM), ACCP, the American Thoracic Society and the Surgical Infection Society. 2 This consensus conference agreed the concept of SIRS, but considered the 1991 definition too non-specific to be useful. The basic definition of sepsis as ‘the clinical syndrome defined by the presence of both infection and a systemic inflammatory response’ remained unchanged but, in place of the SIRS criteria, the 2001 consensus definitions recommend a wider list of ‘possible signs of systemic inflammation in response to infection’. The definitions of severe sepsis as sepsis associated with organ dysfunction, and septic shock as sepsis associated with hypotension despite adequate fluid resuscitation, remained unchanged.

Epidemiology of severe sepsis in the UK NHS

Estimates of severe sepsis in the UK NHS derive from adult critical-care units in the Intensive Care National Audit & Research Centre (ICNARC) Case Mix Programme (CMP) Database. These indicate an increasing treated incidence of severe sepsis in critical care, rising from 50 to 70 cases per 100,000 population per year over the last decade. 3 This now represents approximately 31,000 critical-care unit patient episodes per year. Similarly high incidence rates have been reported elsewhere. 4 Overall, 29% of all admissions to adult, general critical-care units were associated with severe sepsis in the first 24 hours following admission and had an in-hospital mortality of 45%, corresponding to approximately 15,000 deaths per year. These estimates may underestimate the overall burden of severe sepsis within critical-care units in the UK, because of the limitation of the available data restricting analysis to severe sepsis present during the first 24 hours following admission to the critical-care unit.

Severity of severe sepsis has often been summarised by the number of organ dysfunctions (i.e. the number of distinct organ systems with dysfunction). However, although the number of organ dysfunctions is strongly associated with mortality (rising from 22% for single organ dysfunction to 86% for five organ dysfunctions), the particular combination of organ dysfunctions is also important, with the combination of both cardiovascular and renal organ dysfunction associated with particularly high mortality. 5

Intravenous immunoglobulin

Intravenous immunoglobulin (IVIG) is a blood product derived from human donor blood. The serum from around 1000 to 15,000 donors is required for each batch. 6 The mechanisms of action of IVIG are complex, but are increasingly being understood. 7 IVIG is predominantly used in neurology, haematology, immunology and dermatology, but also in nephrology, rheumatology, ophthalmology and other specialties. However, new uses are emerging and off-label use is increasing. 8

Intravenous immunoglobulin has been proposed as an adjuvant therapy for severe sepsis/septic shock since the 1980s and a number of (predominantly small) randomised controlled trials (RCTs) have been conducted. The Cochrane systematic review of the use of IVIG in severe sepsis/septic shock describes the clinical rationale for this as follows: ‘The cascade of harmful effects from sepsis and septic shock has been postulated to be largely due to the lipid A component of the endotoxin molecule in Gram-negative bacteria. Thus the use of antibodies against different components of the endotoxin molecule has been the target of various investigations’. 9 Numerous systematic reviews and meta-analyses of IVIG in severe sepsis/septic shock have been performed. 9–15 As a result of the heterogeneity across studies and inconsistencies in results, the majority of authors have concluded that there is insufficient evidence to recommend IVIG as an adjuvant therapy for severe sepsis/septic shock and that more evidence, in the form of a large, well-conducted RCT, is required.

Current policy and practice with intravenous immunoglobulin for severe sepsis and septic shock in the UK

Intravenous immunoglobulin is a scarce resource worldwide. Costs have escalated, associated with a reduced demand for plasma-derived factor VIII and albumin. In addition, there are supply issues, unique to the UK, that further limit the availability of IVIG. Where IVIG was previously produced in the UK using plasma sourced from within the UK as a by-product of blood donations, plasma must now be imported owing to the risk of variant Creutzfeldt–Jakob disease. In addition, the closure of one UK manufacturer (the Scottish National Blood Transfusion Service) and withdrawal of batches of IVIG because of safety concerns have led to both local and national, transient and longer-term shortages.

In response to this, the Department of Health implemented a Demand Management Programme for IVIG. The programme consists of three components: the Demand Management Plan for Immunoglobulin Use,16 Clinical Guidelines for Immunoglobulin Use17 and the National Immunoglobulin Database. Indications for IVIG use are colour-coded in the following way:

• red – a disease for which treatment is considered the highest priority because of a risk to life without treatment

• blue – a disease for which there is a reasonable evidence base, but where other treatment options are available

• grey – a disease for which the evidence base is weak, in many cases because the disease is rare; treatment should be considered on a case-by-case basis, prioritised against other competing demands, and

• black – a disease for which there is evidence to suggest that IVIG is not an appropriate treatment and treatment is not recommended.

‘Sepsis in the intensive care unit not related to specific toxins or Clostridium difficile’ is currently a black indication and, consequently, IVIG should not be used under any circumstances. The Clinical Guidelines for Immunoglobulin Use do, however, make a research recommendation that, ‘there is a need for adequately powered high-quality RCTs to assess the impact of IVIG in severe sepsis in the general (intensive care unit). 17

In view of the heterogeneity of results from existing RCTs and the unique supply and demand issues for IVIG (especially in the UK), a research priority was identified to establish if such a trial was necessary and feasible and if the costs of carrying out the trial were outweighed by the potential benefit of the resulting information.

Resuscitation bundle

• Measure serum lactate.

• Obtain blood cultures prior to antibiotic administration.

• Administer broad-spectrum antibiotic within 3 emergency department (ED) hours/1 non-ED hour of admission.

• In the event of hypotension and/or serum lactate > 4 mmol/l:

  1. – deliver initial minimum of 20 ml/kg of crystalloid or equivalent

  2. –apply vasopressors for hypotension not responding to initial fluid resuscitation to maintain mean arterial pressure (MAP) ≥ 65 mmHg.

• In the event of persistent hypotension despite fluid resuscitation (septic shock) and/or lactate > 4 mmol/l:

  1. –achieve a central venous pressure (CVP) ≥ 8 mmHg

  2. –achieve a central venous oxygen saturation (ScvO₂) ≥ 70% or mixed venous oxygen saturation (SvO₂) ≥ 65%.

Management bundle

• Administer low-dose steroids for septic shock in accordance with a standardised critical-care policy (if not administered, document why the patient did not qualify for low-dose steroids).

• Administer recombinant human activated protein C (rhAPC) in accordance with a standardised critical-care policy (if not administered, document why the patient did not qualify for rhAPC).

• Maintain glucose control ≥ 3.9 mmol/l, but ≤ 8.3 mmol/l.

• Maintain a median inspiratory plateau pressure < 30 cmH₂O for mechanically ventilated patients.

Statistical analysis

A descriptive analysis was conducted reporting proportion, mean with standard deviation (SD) or median with interquartile range (IQR), as appropriate. Given that for a future RCT of patients with severe sepsis, the recommendation for the control arm would be usual clinical care based on the best available evidence. Adoption of elements from the SSC guidelines ranked level 1A (indicating high-quality evidence and strongly recommended), but which are not included in the resuscitation and management bundles (described above), were examined and reported. These were:

• use of a ventilation weaning protocol

• use of either low-dose unfractionated heparin or low-molecular-weight heparin, unless contraindicated

• use of a mechanical prophylaxis device such as a compression stocking or an intermittent device when heparin is contraindicated

• provision of stress ulcer prophylaxis using an H2 blocker

• contraindicated use of a pulmonary artery catheter (PAC) for routine monitoring of patients with acute lung injury (ALI)/acute respiratory distress syndrome (ARDS).

Components of the resuscitation and management bundles (described above) were also examined and reported. Although not included by the SSC, with strong evidence to support the use of selective decontamination of the digestive tract (SDD), SDD was also examined and reported.

Finally, current use of IVIG was examined and reported.

Survey response

Of the 231 adult, general critical-care units, a senior clinician to complete the survey could not be identified for 14 of the units. Of the remaining 217 units, respondents at four (2%) units refused to complete the survey and completed surveys were received for 123 (57%) units.

Surviving Sepsis Campaign recommendations: level 1A (not included in the bundles)

Responses to the survey for each level 1A item in the SSC guidelines not included in the bundles are reported below.

• Use of a ventilation weaning protocol.

Sixty-three (51%) respondents reported using a ventilation weaning protocol for mechanically ventilated patients in their unit. Overall, respondents estimated that the median proportion of mechanically ventilated patients who were managed using a ventilation weaning protocol was 80% (IQR 50–100%).

• Use of either low-dose unfractionated heparin or low-molecular-weight heparin, unless contraindicated, or a mechanical prophylaxis device such as a compression stocking or an intermittent device when heparin is contraindicated.

All but one of the respondents (n = 122, 99%) reported that they used prophylaxis for deep-vein thrombosis.

• Provision of stress ulcer prophylaxis using an H2 blocker.

All but two of the respondents (n = 121, 98%) reported that they provided stress ulcer prophylaxis.

• Contraindicated use of a PAC for routine monitoring of patients with ALI/ARDS.

A small number of respondents (n = 5, 4%) reported using a PAC.

Resuscitation bundle

The elements that constitute the SSC resuscitation bundle are listed below, along with the strength of the recommendation (1 = strong or 2 = weak) and the quality of evidence (A = high, B = moderate or C = low) assigned by the SSC. 18

• Obtain blood cultures prior to antibiotic administration (1C).

• Administer broad-spectrum antibiotic within 3 ED hours/1 non-ED hour of admission (1B).

• In the event of hypotension and/or serum lactate > 4 mmol/l:

  1. – deliver initial minimum of 20 ml/kg of crystalloid or equivalent (1B)

  2. – apply vasopressors for hypotension not responding to initial fluid resuscitation to maintain MAP ≥ 65 mmHg (1C).

• In the event of persistent hypotension despite fluid resuscitation (septic shock) and/or lactate > 4 mmol/l:

  1. – achieve a CVP of ≥ 8 mmHg (1C)

  2. – achieve an ScvO₂ ≥ 70% or SvO₂ ≥ 65% (1C).

Responses to the survey are reported for each component of the bundle below.

• Obtain blood cultures prior to antibiotic administration (1C).

Nearly all respondents reported that blood cultures are taken in the ED (95%) and in the critical-care unit (98%; Table 1). Respondents estimated that this is carried out for a high proportion of patients presenting at the ED (median 80%, IQR 60–90%) and in almost all patients who are admitted to the critical-care unit (median 100%, IQR 98–100%).

In addition, a high proportion of respondents reported that imaging studies are carried out in the ED and critical-care unit. Although not part of the resuscitation bundle, they are recommended in the SSC guidelines as level 1C (see Table 1).

• Administer broad-spectrum antibiotic within 3 ED hours/1 non-ED hour of admission (1B).

Respondents reported that intravenous antibiotics are given within 1 hour of presentation to the ED (88%) and/or admission to the critical-care unit (93%) (see Table 1). However, they estimated that, on average, a higher proportion of patients receive intravenous antibiotics in the critical-care unit (median 90%, IQR 80–100%) than in the ED (median 60%, IQR 50–80%).

The remaining elements of the resuscitation bundle require specific goals for serum lactate, MAP, CVP and either ScvO₂ or SvO₂. Goals require action that usually translates to the existence of a protocol. Therefore, the survey first asked whether or not the ED and critical-care unit have resuscitation protocols and, if yes, an indication of the clinical parameters included in the protocols.

Forty-one (33%) respondents reported using a resuscitation protocol in the ED and 61 (50%) in the critical-care unit. For the latter, nearly half (n = 29, 48%) of respondents reported that the protocol commenced in the ED and transitioned to the critical-care unit. Although there was variation across hospitals, estimated compliance with the critical care resuscitation protocols was higher (median 77.5%, IQR 60–90%) than with the ED resuscitation protocols (median 60%, IQR 40–70%). The proportions of ED and critical-care unit resuscitation protocols that were reported to include MAP, CVP and ScvO₂/SvO₂ are shown in Figure 2. In addition, respondents reported that ED and critical-care resuscitation protocols also included targets for other parameters that are recommended in the SSC guidelines, but not included in the bundles, e.g. urine output (level 1C) and haemoglobin (level 1B). Nearly all of the critical-care unit resuscitation protocols included targets for cardiac output; however, this was not included in any of the ED resuscitation protocols (see Figure 2).

FIGURE 2.

Reported components of the resuscitation bundle included in ED and critical-care resuscitation protocols.

• In the event of hypotension and/or serum lactate > 4 mmol/l:

  1. – deliver initial minimum of 20 ml/kg of crystalloid or equivalent (1B)

  2. – apply vasopressors for hypotension not responding to initial fluid resuscitation to maintain MAP ≥ 65 mmHg (1C).

Most ED and critical-care resuscitation protocols include serum lactate (see Figure 2) and, although all respondents who answered the question reported aiming to keep serum lactate levels < 4 mmol/l, many reported aiming for ≤ 2 mmol/l.

Both crystalloid and colloid intravenous fluids are used for volume resuscitation; however, respondents reported greater use of crystalloid in the ED than in the critical-care unit, where colloid is used as much as crystalloid (see Table 1).

All respondents reported that MAP is included in the both ED and critical-care unit resuscitation protocols and the majority reported aiming to keep the MAP > 65 mmHg. The reported preferred choice of ‘first-line’ vasopressor in both the ED and critical-care unit was noradrenaline (Figure 3) and the preferred choice of ‘first-line’ inotrope was either dobutamine or adrenaline, although dobutamine was more frequently used in the critical-care unit than in the ED (Figure 4). A small number of respondents (n = 10 and n = 13, respectively) reported that vasopressors and/or inotropes were not given in the ED or were used only with the involvement of critical-care clinicians.

FIGURE 3.

Reported preferred choice of ‘first-line’ vasopressor in the ED and the critical-care unit.

FIGURE 4.

Reported preferred choice of ‘first-line’ inotrope in the ED and the critical-care unit.

• In the event of persistent hypotension despite fluid resuscitation (septic shock) and/or lactate > 4 mmol/l:

  1. – achieve a CVP of ≥ 8 mmHg (1C)

  2. – achieve an ScvO₂ ≥ 70% or SvO₂ ≥ 65% (1C).

Central venous pressure and ScvO₂/SvO₂ were reported less likely to be included in ED than in critical-care resuscitation protocols (see Figure 2). Although there was some variation, most respondents reported aiming for a non-ventilated CVP of ≥ 8 mmHg and a ventilated CVP of around 10–15 mmHg. All respondents reported that they aimed to achieve ScvO₂ of ≥ 70%.

Management bundle

The elements that constitute the management bundle are listed below, along with the strength of the recommendation (1 = strong or 2 = weak) and the quality of evidence (A = high, B = moderate, C = low or D = very low) assigned by the SSC. 18

• Administer low-dose steroids for septic shock in accordance with a standardised critical-care policy (if not administered, document why the patient did not qualify for low-dose steroids) (2C).

• Administer rhAPC in accordance with a standardised critical-care policy (if not administered, document why the patient did not qualify for rhAPC). [2B or 2C for postoperative patients; SSC guidelines state that adult patients with severe sepsis and low risk of death – typically, Acute Physiology and Chronic Health Evaluation (APACHE) II score < 20 or one organ failure – should not receive rhAPC (1A).]

• Maintain glucose control ≥ 3.9 mmol/l but ≤ 8.3 mmol/l (2C).

• Maintain a median inspiratory plateau pressure < 30 cmH₂O for mechanically ventilated patients (1C).

Responses to the survey are reported for each component of the management bundle below.

• Administer low-dose steroids for septic shock in accordance with a standardised critical-care policy (if not administered, document why the patient did not qualify for low-dose steroids) (2C).

A high proportion of respondents (n = 116, 94%) reported that steroids were given in their units for persistent hypotension in septic shock. Although there was variation across units, it was estimated that a high proportion of patients with severe sepsis were given steroids (median 75%, IQR 43–99%).

• Administer rhAPC in accordance with a standardised critical-care policy (if not administered, document why the patient did not qualify for rhAPC). [2B or 2C for postoperative patients; SSC guidelines state that adult patients with severe sepsis and low risk of death – typically, APACHE II score < 20 or one organ failure – should not receive rhAPC (1A).]

A high proportion of respondents (n = 105, 85%) reported that rhAPC was administered to some patients in their unit with severe sepsis. There was variation across units in the proportion of patients who were estimated to receive rhAPC; however, overall, the median proportion was estimated to be 10% (IQR 5–21%).

• Maintain glucose control ≥ 3.9 mmol/l, but ≤ 8.3 mmol/l (2C).

Nearly all respondents (n = 119, 97%) reported that blood glucose control formed part of their unit’s management of patients with severe sepsis. Respondents indicated that blood glucose levels were maintained somewhere within the range of 4–10 mmol/l, although there was variation as to how tightly clinicians aimed to control levels. For example, 35% of respondents reported aiming to keep blood glucose within the range 6–8 mmol/l and 31% within the range 8–10 mmol/l, the latter, higher range resulting from more recent results from a large, multicentre RCT of glucose control. 23

• Maintain a median inspiratory plateau pressure < 30 cmH₂O for mechanically ventilated patients (1C).

Of the 123 respondents, 110 (89%) reported that they aimed to keep the inspiratory plateau pressure < 30 cmH₂O for mechanically ventilated patients. Overall, respondents estimated that this was done for a high proportion of their patients (mean 87.1%, SD 1.4).

Use of selective decontamination of the digestive tract

Only 11 (9%) respondents reported that their unit delivered SDD.

Use of intravenous immunoglobulin

Seventy (56.9%) respondents reported that they used IVIG for advanced management of patients. The clinical reasons given for administering IVIG included neurological diseases, e.g. myasthenia gravis and Guillain–Barré syndrome; toxin-mediated illnesses, e.g. invasive group A streptococcal disease, toxic shock syndrome, necrotising fasciitis, Clostridium difficile colitis, Panton–Valentine leukocidin toxin-producing staphylococcal infection; and other indications, e.g. severe sepsis, liver disease, haematological disease, bronchospasm and immunocompromised patients.

Adoption of resuscitation and management bundles

Overall, 91 (74%) respondents reported that they had adopted a resuscitation bundle and 97 (79%) respondents reported that they had adopted a management bundle. These were mostly the SSC bundles. In addition, 21 respondents reported using the Survive Sepsis UK Sepsis Six24 (Table 2).

Resuscitation

• Take blood cultures.

• Give intravenous antibiotics within 1 hour.

• Maintain serum lactate < 4 mmol/l.

• Fluid resuscitate using a combination of crystalloids and colloids.

• Maintain MAP ≥ 65 mmHg.

• Maintain CVP ≥ 8 mmHg (or 10–15 mmHg for mechanically ventilated patients).

• Give noradrenaline for hypotension not responding to initial fluid resuscitation.

• Maintain ScvO₂ or SvO₂ ≥ 70%.

Management

• Administer low-dose steroids in accordance with standardised critical-care protocol.

• Administer rhAPC in accordance with standardised critical-care protocol.

• Maintain blood glucose levels within the range 4–10 mmol/l.

• Maintain inspiratory plateau pressure < 30 cmH₂O for mechanically ventilated patients.

• Give prophylaxis for deep-vein thrombosis.

• Give stress ulcer prophylaxis.

These results suggest that a protocolised/bundle approach to immediate resuscitation and advanced management would need to be considered for the usual-care arm in any future multicentre RCT of IVIG as an adjunctive therapy in the advanced management of patients acutely ill with severe sepsis. However, specifically with regard to advanced management, a degree of clinical discretion would need to be maintained, illustrated by the high level of variation in compliance with bundle elements in the survey. This variation most likely relates to the heterogeneous nature of the severe sepsis population.

It should be noted that the main limitation of this survey is, despite regular follow-up of non-responders via e-mail and telephone, the low response rate. A major reason for the poor response, based on anecdotal evidence from critical-care clinicians, was the H1N1 swine influenza pandemic. Logistical and management issues took priority over research activities as senior clinicians were required to plan for the pandemic, such as extending critical care areas to be able to cope with additional demands for critical-care services. However, despite the poor response, data from the survey provide useful information on the now widespread adoption, initially resisted, of a protocolised approach to care for patients with severe sepsis in the UK.

Finally, these survey data provide the context for the case mix and outcome data, from the ICNARC CMP Database, used to inform the cost-effectiveness modelling.

Literature searching

The search strategy was divided into four stages.

Inclusion criteria

Inclusion criteria covered design, setting, participants, intervention and outcome measures, as follows:

• design: RCT

• setting: critical-care setting

• participants: adult patients with severe sepsis or septic shock

• intervention: any standard polyclonal IVIG or immunoglobulin (IgM)-enriched polyclonal IVIG (IVIGAM) compared with no intervention, placebo or another standard polyclonal IVIG or IVIGAM preparation

• outcome measures: all-cause mortality, all-cause mortality reported by subgroup and adverse events.

For design, studies that used alternative (rather than randomly generated) allocation sequence were excluded. For participants, studies were included if the majority of patients were aged ≥ 18 years. Clinical judgement was used to determine if the population studied had severe sepsis or septic shock. Studies were assessed by a clinician member of our study team (MSH) and the decision was verified by a clinician member of the Expert Group (MS).

Data extraction

Data were extracted from studies by two independent reviewers (NJW and JJM) using a standardised data extraction spreadsheet. Duplicate extraction was performed for 9/17 (53%) of the studies and any differences were resolved through discussion. Extracted data from all studies were compared with extracted data reported in the previous systematic reviews identified from stage 1 of the literature searching (see Stage 1: previous systematic reviews). Finally, all clinical data were double-extracted by a clinician on the study team (MSH) and any queries addressed and confirmed through discussion with clinical experts on the Expert Group (WACS and MS).

Data extraction covered details, quality, population, intervention and outcomes for each study. Data were extracted for all, where available.

• Details: date recruitment started; study duration; publication date; critical-care setting reported; whether or not multicentre and, if so, the number of centres.

• Quality: whether or not concealment of allocation to treatment was adequate/unclear/inadequate; whether or not blinding to treatment was adequate/unclear/inadequate; whether or not randomisation procedure was adequate/unclear/inadequate; whether or not an intention-to-treat analysis was performed; whether or not the trial received funding from industry sponsors; and the Jadad score,25 which is based on a composite score for adequacy of randomisation (0–2 points), blinding (0–2 points) and presence or absence of attrition information (0–1 points), yielding a score from 0 to 5, where 5 represents the best-quality score.

• Population: study inclusion and exclusion criteria; proportion of male/female patients; mean age; proportion of patients with septic shock; baseline severity scores [APACHE II score26; Simplified Acute Physiology Score (SAPS) II27; Sequential Organ Failure Assessment (SOFA)28; Sepsis Score29]; multiorgan dysfunction/organ failure/number and type of organ failures.

• Intervention: IVIG product used in the intervention arm(s); information on dosing, including daily dose (g/kg day); volume of fluid given (ml/kg day); duration of treatment (days); total dose (g/kg); description of the control intervention.

• Outcomes: number of events (deaths) out of total number of patients per trial arm; follow-up duration; any reported adverse events; duration of mechanical ventilation; duration of critical-care unit stay; and duration of acute hospital stay.

Data analysis

Data analysis was divided into descriptive analyses and modelling.

Literature searching/inclusion criteria

Data extraction

The data extracted from the 17 included studies42–55 are summarised in Tables 7–13.

Study quality and publication bias

Table 8 reports the assessment of study quality metrics. Concealment of allocation to treatments was considered adequate in 5/17 (29.4%) studies,42,43,47,53,58 inadequate in 2/17 (11.8%) studies,52,57 and it was not declared and was therefore unclear from the published paper for the remaining 10/17 (58.8%) studies. 44–46,48–51,54–56 For blinding of patients and assessors to treatment received was considered adequate in 5/17 (29.4%) studies,42,50,51,53,56 inadequate in 5/17 (29.4%) studies,43,47,52,57,58 and it was unclear in the published paper for the remaining 7/17 (41.2%) studies. 44–46,48,49,54,55 The majority of the studies, 9/17 (52.9%),42,43,45,47–49,53,54,58 used an appropriate method of randomisation, although this was unclear in the published paper in the remaining 8/17 (47.1%) studies. 44,46,50–52,55–57 An intention-to-treat analysis was performed in 12/17 (70.6%) studies,42,43,45,46,48,49,52–56,58 was not performed in 3/17 (17.6%) studies,47,50,51 and it was unclear from the published paper for the remaining 2/17 (11.8%) studies. 44,57 The Jadad score, a composite measure of study quality ranging from 0 to 5 (where 5 represents best study quality). This analysis revealed that only 4/17 (23.5%) studies achieved a Jadad score of 5,42,51,53,56 7/17 (41.2%) achieved a Jadad score of 3,43,45,47,50,55,57,58 and the remaining 6/17 (35.3%) studies achieved a Jadad score of ≤ 2. 44,46,48,49,52,54 Industry sponsorship was acknowledged in 7/17 (41.2%) studies. 42,43,47,51,53,56,58 For the remaining studies it was unclear if there was industry sponsorship. 44–46,48–50,52,54,55,57

TABLE 8 - Study quality

Study number	Study	Allocation concealment	Blinding	Randomisation	Intention-to-treat analysis?	Jadad score	Industry sponsorship declared?
1	Rodriguez et al. (2005)42	Adequate	Adequate	Adequate	Yes	5	Yes
2	Hentrich et al. (2006)43	Adequate	Inadequate	Adequate	Yes	3	Yes
3	Karatzas et al. (2002)44	Unclear	Unclear	Unclear	Unclear	2	No
4	Tugrul et al. (2002)45	Unclear	Unclear	Adequate	Yes	3	No
5	Behre et al. (1995)46	Unclear	Unclear	Unclear	Yes	1	No
6	Schedel et al. (1991)47	Adequate	Inadequate	Adequate	No	3	Yes
7	Wesoly et al. (1990)48	Unclear	Unclear	Adequate	Yes	1	No
8	Spannbruker et al. (1987)49	Unclear	Unclear	Adequate	Yes	1	No
9	Dominioni et al. (1996)50	Unclear	Adequate	Unclear	No	3	No
10	Burns et al. (1991)51	Unclear	Adequate	Unclear	No	5	Yes
11	De Simone et al. (1988)52	Inadequate	Inadequate	Unclear	Yes	1	No
12	Werdan et al. (2007)53	Adequate	Adequate	Adequate	Yes	5	Yes
13	Grundmann and Hornung (1988)54	Unclear	Unclear	Adequate	Yes	2	No
14	Darenberg et al. (2003)56	Unclear	Adequate	Unclear	Yes	5	Yes
15	Lindquist et al. (1981)57	Inadequate	Inadequate	Unclear	Unclear	3	No
16	Masaoka et al. (2000)58	Adequate	Inadequate	Adequate	Yes	3	Yes
17	Yakut et al. (1998)55	Unclear	Unclear	Unclear	Yes	3	No

Figure 5 shows a funnel plot of the standard error (SE) of the effect size (log-OR) plotted against study effect size (OR on the log-scale). From this plot, it can be seen that there does appear to be funnel-plot asymmetry, where there are studies ‘missing’ from the right-hand-side of the plot when the SE is high (bottom right of plot), and this is supported by the Peters test for small-study effects (p = 0.0008). This suggests that there may potentially be an issue with publication bias with this evidence.

FIGURE 5.

Publication bias – funnel plot (with pseudo 95% confidence limits) of mortality for IVIG and IVGAM versus control.

Baseline patient characteristics

Participants in the studies are described in Table 9. The baseline patient characteristics were used to identify if studies met the severe sepsis/septic shock eligibility criterion for inclusion in our review.

TABLE 9 - Details of participants (severe sepsis/septic shock)

STSS, streptococcal toxic shock syndrome.

Study number	Study	Sepsis definition	Additional enrolment criteria and definitions
1	Rodriguez et al. (2005)42	ACCP/SCCM criteria	Severe sepsis/septic shock of intra-abdominal origin admitted to a critical-care unit within 24 hours of onset of symptoms. Abdominal sepsis defined by the presence of SIRS and a surgically confirmed abdominal focus. Obtaining purulent material or detecting potential pathogens using Gram staining was mandatory. Appropriateness of the surgical procedure (successful eradication of focus), according to criteria of the attending surgical team and the intensivist, required for inclusion
2	Hentrich et al. (2006)43	ACCP/SCCM criteria	Sepsis syndrome and: diagnosis of haematological malignancy neutropenia
3	Karatzas et al. (2002)44	ACCP/SCCM criteria	Severe sepsis
4	Tugrul et al. (2002)45	ACCP/SCCM criteria	Severe sepsis
5	Behre et al. (1995)46	ACCP/SCCM criteria	Sepsis syndrome and: diagnosis of haematological malignancy neutropenia
6	Schedel et al. (1991)47	‘Septic shock’	Detection of endotocaemia (> 12.5 pg/ml endotoxin) and at least five of the following criteria: clinical indications of septicaemia fever ≥ 38.5°C platelet count < 100 x 109/l or a 30% drop in the last 24 hours shift to left in the blood count granulocytopenia pulmonary congestion disseminated intravascular coagulation systolic blood pressure < 100 mmHg heart rate > 120/min urine output < 500 ml/day
7	Wesoly et al. (1990)48	Sepsis score ≥ 12	Postoperative
8	Spannbruker et al. (1987)49	‘Septic shock’
9	Dominioni et al. (1996)50	Sepsis score ≥ 17	Sepsis following surgery or trauma
10	Burns et al. (1991)51	Platelet count < 75 × 109/l Documentation of suspected infection with positive culture	Suspected infection documented by one or more of the following: fever leukocytosis elevated band neutrophil count infiltrate on radiography of chest consistent with pneumonia toxic granulations or Döhle bodies on peripheral smear positive Gram stain of body fluid or exudates
11	De Simone et al. (1988)52	‘Severe sepsis’
12	Werdan et al. (2007)53	At least four of nine ‘sepsis criteria’	(1) Sepsis criteria: temperature > 38.5°C or < 36°C white blood cell count > 12 × 109/l or < 3.5 × 109/l heart rate > 100/min respiratory rate > 28/min or fraction of inspired oxygen (FiO2) > 0.21 mean arterial pressure < 75 mmHg cardiac index > 4.5 l/min/m or systemic vascular resistance < 800 dyn/s/cm platelet count < 100 × 109/l positive blood cultures clinical evidence of sepsis (surgical or invasive procedure during the preceding 48 hours or presence of an obvious septic focus) (2) Sepsis score 12–27 (3) APACHE II score 20–35
13	Grundmann and Hornung (1988)54	Sepsis score > 12	Postoperative Gram-negative bacterial infection with positive endotoxin in plasma for 2 subsequent days
14	Darenberg et al. (2003)56	STSS consensus definition	Patients could be enrolled before results from bacteriological cultures were obtained if they had clinical symptoms of STSS and if a streptococcal infection was suspected
15	Lindquist et al. (1981)57	Sepsis secondary to ‘septicaemia’ based on Svanbom criteria	Purulent meningitis irrespective of aetiology Suspected or verified bacterial pneumonia (day-time admissions only)
16	Masaoka et al. (2000)58	ACCP/SCCM criteria	Suspected sepsis, as defined by heart rate > 90/min, respiratory rate > 20/min, in addition to positive C-reactive protein and sustained fever ≥ 38°C with: specific infection, e.g. respiratory tract infection such as pneumonia, urinary tract infection no tumour, transfusion, drug-induced fever blood culture-negative Patients were randomised if they were ‘non-responders’ – did not have enough improvement of symptoms with administration of broad-spectrum antibiotics for more than 3 consecutive days (72 hours)
17	Yakut et al. (1998)55	Sepsis score > 16	Post-surgical

Summary patient baseline characteristics are reported in Table 10. Mean age was broadly comparable across treatment arms, both within and across studies. Mean severity was broadly comparable across treatment arms within studies but differed between studies, highlighting the heterogeneity in the severity of the severe sepsis/septic shock patients recruited into the different studies. The proportion of male patients randomised varied not just across studies, but also between treatment arms within studies. None of the studies reported mortality separately for men and women and, so, it is not possible to assess whether or not this baseline imbalance might introduce bias in the results. Similarly, where reported,42,43,46,52,53 the proportion of patients randomised with septic shock, rather than other severe sepsis, differed both across studies and across treatment arms within studies. Rodriguez et al. ,42 Hentrich et al. 43 and Behre et al. 46 reported results differentially by septic shock or other severe sepsis. These results showed that mortality rates were much lower for patients with septic shock than for those with severe sepsis; however, the treatment effects within these two subgroups did not differ substantially (Figure 6).

TABLE 10 - Baseline patient characteristics

NR, not reported.

Approximated from figure.

Description of sepsis syndromes provided suggests high severity of illness.

Study number	Study	Age: mean (SD), years		Severity of illness: measure, mean (SD)
		IVIG	Control	IVIG	Control
1	Rodriguez et al. (2005)42	61.3 (19.9)	65.9 (18.2)	APACHE II 16.1 (5.9)	APACHE II 15.2 (6.1)
2	Hentrich et al. (2006)43	48.8 (NR)	51.0 (NR)	NR	NR
3	Karatzas et al. (2002)44	50.5 (3.33)	50.7 (7.4)	APACHE II 21.3 (7.2)	APACHE II 23.5 (7.9)
4	Tugrul et al. (2002)45	42 (18)	49.3 (20.6)	APACHE II 10.5 (4.6)	APACHE II 14 (8.5)
5	Behre et al. (1995)46	50 (NR)	55 (NR)
6	Schedel et al. (1991)47	46 (16)	37 (18)	APACHE II 30a	APACHE II 24a
7	Wesoly et al. (1990)48	44.7 (19)	54.8 (17)	Sepsis score 14.8 (2.5)	Sepsis score 16.3 (3.6)
8	Spannbruker et al. (1987)49	50.8 (15.5)	54.5 (12)	NR	NR
9	Dominioni et al. (1996)50	55 (19)	57 (19)	Sepsis score 23 (4)	Sepsis score 23 (4)
10	Burns et al. (1991)51	61.5 (NR)	59.8 (NR)	NR	NR
11	De Simone et al. (1988)52	45 (4)	45 (5)	NRb	NRb
12	Werdan et al. (2007)53	57.2 (13.7)	57.7 (13.6)	APACHE II 27.6 (4.5)	APACHE II 28 (4.5)
13	Grundmann and Hornung (1988)54	46.9 (NR)	52.8 (NR)	NR	NR
14	Darenberg et al. (2003)56	51.3 (NR)	52.6 (NR)	SAPS II 53 (NR) SOFA 11 (NR)	SAPS II 51 (NR) SOFA 11 (NR)
15	Lindquist et al. (1981)57	48.3 (NR)	39.2 (NR)	NR	NR
16	Masaoka et al. (2000)58	NR	NR	NR	NR
17	Yakut et al. (1998)55	32 (16)	31 (16)	APACHE II 16 (4)	APACHE II 16 (5)

FIGURE 6.

Forest plot for fixed-effects model using inverse variance weights – IVIG/IVIGAM versus control for those studies reporting results by whether the patients have septic shock or severe sepsis.

Interventions

Table 11 describes the preparations used for the control and IVIG arms with the dosing regimes reported. In all cases, the control and IVIG arms were given as adjunct therapy to standard care, although standard care varied between studies. In 8/17 (47.1%) of the studies,42–49 Pentaglobin^® (IVIGAM, Biotest Pharma, Germany), an IVIGAM, was used. In all other studies, standard preparations of IVIG were used; however, variation existed in the standard IVIG preparations used. The dosing regimes used also varied between the studies. Duration of treatment ranged from 2 days to 7 days (for some of the longer durations, treatment was not given on every day), with the majority of studies43–49,51,56–58 [11/17 (64.7%)] using a duration of 3 days. Average daily dose ranged from 0.07 g/kg/day to 0.67 g/kg/day, volume given ranged from 1.4 ml/kg/day to 13.34 ml/kg/day and total dose ranged from 0.45 g/kg to 2 g/kg. These dosing variables were inter-related as presented in Figure 7. There was a near perfect relationship between volume and average daily dose, which arose because nearly all studies used a 5% preparation. There was a negative relationship between volume (and average daily dose) and duration. These relationships may reflect the differences in dosing recommendations between different underlying disease conditions and the severity of sepsis.

TABLE 11 - Description of interventions and dosing regimes

HAS, human albumin solution.

Pentaglobin is IVIGAM; all other preparations are standard IVIG.

Where not reported, a 5% preparation assumed and used to calculate daily dose, volume and total dose; if dose was not given as per kg body weight, then a typical body weight of 70 kg assumed to obtain per kg body weight.

Study	Control	IVIG preparationa	IVIG dosing regime	Average daily doseb (g/kg/day)	Volume (ml/kg/day)	Duration of therapy (days)	Total dose (g/kg)
Rodriguez et al. (2005)42	5% HAS	Pentaglobin (Biotest Pharma, Germany)	0.35 g/kg/day	0.35	7	5	1.75
Hentrich et al. (2006)43	HAS	Pentaglobin (Biotest Pharma, Germany)	1300 ml over 72 hours: 200 ml initially (0.5 ml/min) then 11 infusions of 100 ml every 6 hours	0.31	6.2	3	0.93
Karatzas et al. (2002)44	No treatment	Pentaglobin (Biotest Pharma, Germany)	5 ml kg/day over 6 hours	0.25	5	3	0.75
Tugrul et al. (2002)45	No treatment	Pentaglobin (Biotest Pharma, Germany)	5 ml kg/day over 6 hours	0.25	5	3	0.75
Behre et al. (1995)46	5% HAS	Pentaglobin (Biotest Pharma, Germany)	Loading dose 10 g then 5 g 6-hourly for 72 hours	0.31	6.2	3	0.93
Schedel et al. (1991)47	No treatment	Pentaglobin (Biotest Pharma, Germany)	Loading dose 600 ml over 8 hours then two further doses of 300 ml every 24 hours	0.285	5.7	3	0.855
Wesoly et al. (1990)48	No treatment	Pentaglobin (Biotest Pharma, Germany)	0.25 g/kg/day	0.25	5	3	0.75
Spannbruker et al. (1987)49	No treatment	Pentaglobin (Biotest Pharma, Germany)	0.15 g/kg/day	0.15	3	3	0.45
Dominioni et al. (1996)50	5% HAS	Sandoglobulin (Sandoz Pharmaceutical Corp, Italy)	0.4 g/kg on day 0 0.4 g/kg 24 hours later 0.2 g/kg 5 days later	0.2	4	5	1
Burns et al. (1991)51	HAS	Sandoglobulin (Sandoz Pharmaceutical Corp, Italy)	0.4 g/kg/day	0.4	8	3	1.2
De Simone et al. (1988)52	No treatment	Sandoglobulin (Sandoz Pharmaceutical Corp, Italy)	0.4 g/kg on day 0 0.2 g/kg 48 hours later 0.4 g/kg 5 days later	0.2	3.33	5	1
Werdan et al. (2007)53	0.1% HAS	Polyglobin N (Bayer Biological Products, Germany)	0.6 g/kg on day 0 0.3 g/kg on day 1 or 2	0.45	9	2	0.9
Grundmann and Hornung (1988)54	No treatment	Intraglobin F (Biotest Pharma, Germany)	0.25 g/kg/day	0.25	5	2	0.5
Darenberg et al. (2003)56	1% HAS	Endobulin SD (Baxter)	Loading dose of 1 g/kg then 0.5 g/kg every 24 hours for three doses	0.667	13.34	3	2.001
Lindquist et al. (1981)57	No treatment	Pepsin-treated human gamma globulin – Gamma-venin	0.15 g/kg over 1 hour	0.15	3	3	0.45
Masaoka et al. (2000)58	No treatment	Not specified	5 g/day for 3 consecutive days	0.07	1.4	3	0.21
Yakut et al. (1998)55	20% HAS	Gamimune N 10% (Miles Inc. Pharmaceutical Division, USA)	0.4 g/kg on day 0 0.4 g/kg on day 1 0.2 g/kg on days 2–7	0.26	5.2	7	1.8

FIGURE 7.

Relationships between volume, average daily dose and duration of treatment.

For the analyses, several ways to allow for differences between the treatments and dosing regimes were considered. For the different IVIG and control preparations, five different possible treatment comparison models (numbered according to number of treatments) were considered:

• model T2 – IVIG or IVIGAM versus albumin or no treatment

• model T3a – IVIG versus IVIGAM versus albumin or no treatment

• model T3b – IVIG or IVIGAM versus albumin versus no treatment

• model T4 – IVIG versus IVIGAM versus albumin versus no treatment

• model T10 – Sandoglobin^® versus Intraglobin versus Gamma-Venin versus Polyglobin versus Endobulin versus Gamumin N versus IVIG unspecified versus IVIGAM versus albumin versus no treatment.

For the dosing regimes, extending the range of treatment comparison models according to dose was considered, but these models did not always result in a connected network of treatment comparisons. For those models that could be fitted, there was little to be gained from this approach. Dosing regime had multiple attributes and it was not clear how to define the treatments in this way. Instead, the attributes of the dosing regime (average daily dose, volume, duration and total dose) were considered as covariates for the five treatment comparison models described above.

Data analysis

Descriptive analyses

For all of the descriptive analyses, treatment model T2, comparing IVIG/IVIGAM versus albumin/no treatment, was used. All treatment effects are displayed as ORs with 95% confidence intervals (CIs). Figures 8 and 9 present forest plots for a fixed- and a random-effects meta-analysis, respectively. There is evidence of heterogeneity in the treatment effects (I² = 46.9%, Q = 30.1, df = 16, p = 0.017). The pooled OR from the fixed-effects model is 0.68 (95% CI 0.54 to 0.84), showing a reduction in the odds of mortality with IVIG/IVIGAM compared with albumin/no treatment. The pooled OR from the random-effects model is 0.47 (95% CI 0.32 to 0.69), showing a stronger effect. Note that the large weight of the Werdan et al. 53 study drives the difference between the fixed- and random-effects models’ results because it is given less weight in the random-effects model.

FIGURE 8.

Forest plot for fixed-effects model using inverse variance weights – IVIG and IVIGAM treatments versus control.

FIGURE 9.

Forest plot for random-effects model using inverse variance weights – IVIG and IVIGAM versus control.

Heterogeneity in the study results was further explored by looking at the descriptive results for different values/subgroups of potential explanatory factors. In all cases, fixed-effects models are reported. Where the explanatory variable was a continuous measure (e.g. daily dose), studies were grouped into quartiles to explore trends in treatment effect over the continuous measure.

Type of intravenous immunoglobulin and control treatment

Figure 10 indicates that there is a slightly stronger effect for studies that used an enriched IVIGAM product (OR 0.54, 95% CI 0.37 to 0.79) than for studies that used a standard polyclonal IVIG (OR 0.76, 95% CI 0.58 to 0.99). However, this difference did not explain a large amount of the heterogeneity observed (p = 0.156 for heterogeneity between groups). Choice of control (albumin or no treatment) had a strong influence on the pooled treatment effect (Figure 11), with studies using albumin as control giving a pooled OR of 0.80 (95% CI 0.62 to 1.02) compared with an OR of 0.36 (95% CI 0.22 to 0.58) in studies that used no treatment as control. The choice of control explained some of the heterogeneity between studies (p = 0.004 for heterogeneity between groups). Possible reasons for this may relate to the fact that use of albumin introduces an intervention that may have biological effects. Two possible options are (1) albumin is an effective treatment for severe sepsis68 or (2) the use of albumin makes the control treatment appear similar to the IVIG treatment (i.e. slightly frothy) and the use of albumin as control is simply an indicator of appropriate blinding in these studies and could, therefore, possibly be a proxy for the risk of bias being lower in these studies.

FIGURE 10.

Fixed-effects model by IVIG preparation (IVIG or IVIGAM).

FIGURE 11.

Fixed-effects model by choice of control (no treatment or albumin).

Dosing regimes

The pooled treatment effect becomes stronger (OR decreases) with increasing duration of treatment (p = 0.001 for heterogeneity between groups; Figure 12), becomes less strong (OR increases) with increasing daily dose and volume (p = 0.001 for heterogeneity between groups; Figures 13 and 14) and shows no clear pattern with total dose (Figure 15). Specific aspects of the dosing regime in the studies appeared to have a strong explanatory effect on the observed treatment effect. Whether the dosing regime itself was the cause of this difference or it was simply a measure that was confounded with other differences between the studies could not be determined from the available evidence. In particular, the absence of any dose-finding studies for IVIG preparations prevented us from drawing conclusions about the effect of dosing regime on treatment effect.

FIGURE 12.

Fixed-effects model by duration of treatment (days).

FIGURE 13.

Fixed-effects model by quartiles of daily dose (g/kg/day).

FIGURE 14.

Fixed-effects model by quartile of volume (ml/kg/day).

FIGURE 15.

Fixed-effects model by quartiles of total dose (g/kg).

Study quality

Study quality indicators that were explored included intention-to-treat analysis, concealment of allocation, blinding to treatment and randomisation procedure.

In all cases (Figures 16–19), the pooled treatment effect was less strong (OR increased) when the study was considered adequate on each indicator. These study quality indicators explained a large amount of heterogeneity (p < 0.02 for heterogeneity between groups in all cases except intention-to-treat analysis, where p = 0.048).

FIGURE 16.

Fixed-effects model by whether or not intention-to-treat analysis performed.

FIGURE 17.

Fixed-effects model by adequacy of concealment of allocation to treatment.

FIGURE 18.

Fixed-effects model by adequacy of blinding to treatment.

FIGURE 19.

Fixed-effects model by adequacy of randomisation procedure.

The Jadad score is a composite measure of study quality, and the pooled treatment effect was less strong (OR closer to 1) when the Jadad score was 5 (best quality score), compared with lower scores (Figure 20). Jadad score explained a large amount of the heterogeneity between studies (p = 0.004 for heterogeneity between groups).

FIGURE 20.

Fixed-effects model by Jadad score (where 5 represents the best quality).

Studies that acknowledged industry sponsorship showed less strong treatment effect (OR closer to 1) than studies that did not acknowledge industry sponsorship (Figure 21) and this explained a large amount of heterogeneity (p < 0.001 for heterogeneity between groups). However, this result is dominated by the two very large studies41,51 acknowledging industry sponsorship. Because it was not clear if those studies that did not acknowledge industry sponsorship were sponsored or not (the information was essentially missing), the interpretation of this covariate is difficult and we, therefore, did not include this covariate in the modelling exercise.

FIGURE 21.

Fixed-effects model by whether industry sponsorship reported.

The pooled treatment effect was less strong (OR closer to 1) for studies published more recently than for older studies (Figure 22). This explained a large amount of heterogeneity (p = 0.001 for heterogeneity between groups). There was a trend across studies by quartile of sample size (Figure 23), indicating that pooled treatment effects were stronger (OR smaller) for studies with smaller sample sizes than for studies with larger sample sizes (p < 0.001 for heterogeneity between groups). Figure 24 presents the same pattern with sample size (N), but plotted for 1/N which is a useful way to model small-study effects. 69,70 This is because, as N gets large, 1/N becomes small, so the results for 1/N=0 can be interpreted as an effect estimate for ‘very large’ studies, i.e. adjusted for small study effects.

FIGURE 22.

Fixed-effects model by quartile of publication date.

FIGURE 23.

Fixed-effects model by quartile of sample size for IVIG arm (first quartile = smallest to fourth quartile = largest).

FIGURE 24.

Fixed-effects model by quartile of 1/N (first quartile = smallest to fourth quartile = largest).

In summary, the exploration of study quality indicators identified that there appeared to be issues with study quality, potential publication bias and other small-study effects in the available evidence.

Other factors

Whether the study was clearly conducted in a critical-care setting, or not, did not have any effect on the pooled treatment effect (Figure 25). This was not surprising as our inclusion criteria limited selection to those studies that were deemed to be in a severe sepsis/septic shock population. It is highly likely that all of studies were conducted in a critical-care setting irrespective of whether or not this was reported.

FIGURE 25.

Fixed-effects model by whether or not the study clearly took place in a critical-care setting.

Figures 26 and 27 indicate that the pooled treatment effect was stronger (OR smaller) in studies conducted in populations with a higher baseline risk (third and fourth quartiles) than in studies conducted in populations with lower baseline risk (first and second quartiles). This explained a large amount of heterogeneity (p < 0.001 for heterogeneity between groups).

FIGURE 26.

Fixed-effects model by quartile of baseline risk (control arm log-odds of mortality) (first quartile = lowest risk to fourth quartile = highest risk).

FIGURE 27.

Fixed-effects model by pooled quartiles of baseline risk (control arm log-odds of mortality) (first and second quartiles = low risk and third and fourth quartiles = high risk).

There was no clear pattern between the pooled treatment effect and the follow-up period used by the studies (Figure 28). Although there was clearly a relationship between mortality and follow-up period (see Table 12), the relative effects do not appear to depend strongly on follow-up period.

FIGURE 28.

Fixed-effects model by follow-up period.

Modelling

Two-treatment comparison model – all-cause mortality

Table 14 presents model fit results from fixed- and random-effects models with no covariates for the two-treatment comparison model (model T2: IVIG or IVIGAM vs albumin or no treatment). For a well-fitting model, the posterior mean residual deviance, D¯res, would be expected to be approximately equal to the number of data points (e.g. 34 from 17 studies each with two arms). Models in which D¯res is much larger than this display evidence of lack of fit. The DIC provides a composite measure of model fit and complexity and models are preferred with lower DIC. Differences in D¯res and DIC of ≤ 2 are considered meaningful. 37 The fixed-effects model shows substantial lack of fit (D¯res=51.4), whereas the random-effects model fits well (D¯res=30.9). This highlights the heterogeneity present in the available evidence. The random-effects model is to be preferred on the basis of both model fit (D¯res) and DIC.

TABLE 14 - Summaries of model fit for the two-treatment comparison model [model T2: (IVIG/IVIGAM) vs (albumin/no treatment)] with different key covariates

For a good-fitting model, a D¯res approximately equal to 34 (the number of data points) would be expected and values much greater than this suggest lack of fit.

The DIC is a composite measure of fit and complexity models with smaller DICs (where differences of ≥ 2 are considered meaningful differences) are preferred.

Model	Posterior mean residual deviance, D¯resa	DICb	Posterior mean between trials heterogeneity (SD), τ¯
No covariates
Random-effects model	30.9	175.0	0.56
Fixed-effects model	51.4	188.2
Fixed-effects model adding covariates (individually) for dosing regime
Duration of treatment (days)	37.1	175.0
Daily dose (g/kg/day)	36.9	174.6
Volume (ml/kg/day)	36.9	174.6
Total dose (g/kg)	52.2	190.0
Fixed-effects model adding covariates (individually) for study quality
Whether or not intention-to-treat analysis performed	45.0	182.7
Adequacy of concealment of allocation to treatment	41.5	179.2
Adequacy of blinding to treatment	48.8	186.5
Adequacy of randomisation procedure	45.2	182.9
Jadad score	39.2	176.9
Publication date	35.9	173.7
1/N (N = number of patients randomised to the IVIG arm)	36.6	174.4
Fixed-effects model adding covariates (individually) for other factors
Critical-care setting	51.6	189.4
Baseline risk (control arm log-odds of mortality)	53.0	190.8
Follow-up period (linear relationship)	46.5	184.3
Follow-up period (< 4 weeks or ≥ 4 weeks)	48.5	186.3
Fixed-effects model adding combinations of key covariates (i.e. results in bold above)
Duration of treatment + daily dose + volume	34.3	173.6
Jadad score + publication date + 1/N	35.7	175.4
Duration of treatment + Jadad score	33.4	172.3
Duration of treatment + publication date	31.4	170.2
Duration of treatment + 1/N	33.7	172.5
Daily dose + Jadad score	37.4	176.2
Daily dose + publication date	34.6	173.3
Daily dose + 1/N	32.2	171.0
Volume + Jadad score	37.5	176.3
Volume + publication date	34.7	173.4
Volume + 1/N	32.4	171.2
Random-effects model adding key covariates (individually) (i.e. results in bold above)
Duration of treatment (days)	32.5	175.3	0.38
Daily dose (g/kg/day)	33.0	175.2	0.36
Volume (ml/kg/day)	33.2	175.3	0.36
Jadad score	32.2	175.6	0.45
Publication date	33.1	174.8	0.31
1/N (N = number of patients randomised to the IVIG arm)	32.7	174.6	0.33

Table 14 also presents model fit statistics for the fixed-effects model for the two-treatment comparison model (Model T2: IVIG or IVIGAM vs albumin or no treatment) with a range of covariates included individually (i.e. univariate analyses). Results for covariates that substantially improve model fit are highlighted in bold (see Table 14). The key covariates that appeared to explain the heterogeneity in these studies were dosing regime covariates [duration of treatment (days), daily dose (g/kg/day) and volume (ml/kg/day)] and study quality covariates (Jadad score, publication date and a measure of sample size; 1/N). Including any one of these key covariates resulted in a model that fitted adequately and, on the basis of the DIC, was comparable with the random-effects model with no covariates. In other words, these key covariates explained the majority of the heterogeneity (as indicated by the reduction in the posterior mean between study SD, τ¯, when the random-effects models was fitted with these covariates). Follow-up period showed a mild effect, but this disappeared when any of the above key covariates were included (results not shown). Including all three dosing regime covariates did not improve model fit (see Table 14) and, therefore, the conclusion was that, as long as one of these three aspects of dosing regime was included, it was not necessary to include the other two. In other words, these three dosing regime covariates were explaining the same aspects of heterogeneity in treatment effect across the studies. Similar results were observed for the three key study quality covariates and it was only considered necessary to include one of these three covariates in further models.

Combining a dosing regime covariate with a study quality covariate improved model fit and led to reductions in DIC (see Table 14). This suggested that these two types of covariates were both measuring different aspects of heterogeneity across the studies. The fixed-effects models that gave the lowest DIC are highlighted in bold in Table 14 and listed below:

• duration of treatment + Jadad score

• duration of treatment + publication date

• duration of treatment + 1/N

• daily dose + 1/N

• volume + 1/N.

Discussions with the Expert Group highlighted that there was no clear clinical rationale why duration of treatment, daily dose or volume would affect treatment efficacy or effectiveness. For this reason, random-effects models with solely study quality covariates were considered. The heterogeneity that can be explained with the dosing regime covariates was left unexplained in these models, reflecting a belief that these covariates were a proxy for other, unmeasured, differences between the studies.

Comparing treatment comparison models T2, T3a, T3b, T4 and T10

Table 15 presents model fit summaries for the key covariates for the treatment comparison models T2, T3a, T3b, T4 and T10. For the fixed-effects model with no covariates, model fit was improved for models T3b and T4 that included albumin and no treatment as separate treatments, compared with models T2 and T3a that did not distinguish between the treatments given in the control arm. Further improvement in model fit was seen by treating each IVIG preparation as a separate treatment (treatment comparison model T10); however, the DIC was higher for this model than for the treatment comparison models T3a and T3b owing to the increased complexity (number of parameters). On the basis of DIC, model T3b (IVIG or IVIGAM vs albumin vs no treatment) was preferred as the model providing the most parsimonious compromise between model fit and complexity.

TABLE 15 - Summaries of model fit for models including key covariates (identified from Table 12) for the range of treatment comparison models

For a good-fitting model, a D¯res approximately equal to 34 (the number of data points) would be expected and values much greater than this suggest lack of fit.

The DIC is a composite measure of fit and complexity models with smaller DICs (where differences of ≥ 2 are considered meaningful differences) are preferred.

Model	T2		T3a		T3b		T4		T10
	D¯resa	DICb	D¯resa	DICb	D¯resa	DICb	D¯resa	DICb	D¯resa	DICb
No covariates
Random-effects model	30.9	175.0			31.6	175.3
Fixed-effects model	51.4	188.2	50.1	187.9	42.8	180.5	43.6	182.3	36.2	180.7
Fixed-effects model adding covariates (individually) for dosing regime
Duration of treatment (days)	37.1	175.0	37.6	176.4	29.8	168.6	30.7	170.5	34.5	180.1
Daily dose (g/kg/day)	36.9	174.6	38.0	176.7	37.4	176.2	38.3	178.1	36.5	182.1
Volume (ml/kg/day)	36.9	174.6	38.0	176.8	37.5	176.3	38.4	178.3	36.5	181.9
Fixed-effects model adding covariates (individually) for study quality
Jadad score	39.2	176.9	39.7	178.4	39.3	178.1	39.7	179.5	37.1	182.6
Publication date	35.9	173.7	36.5	175.3	36.4	175.2	37.2	179.5	36.3	182.6
1/N	36.6	174.4	37.3	176.0	36.1	174.9	36.8	176.6	37.3	182.9
Fixed-effects model adding combinations of key covariates
Duration of treatment + Jadad score	33.4	172.3			30.7	170.5
Duration of treatment + publication date	31.4	170.2			30.1	169.8
Duration of treatment + 1/N	33.7	172.5			30.7	170.5
Daily dose + Jadad score	37.4	176.2			38.0	177.7
Daily dose + publication date	34.6	173.3			35.6	175.4
Daily dose + 1/N	32.2	171.0			33.1	172.9
Volume + Jadad score	37.5	176.3			38.1	177.8
Volume + publication date	34.7	173.4			35.7	175.5
Volume + 1/N	32.4	171.2			33.3	173.1

Including covariates, the best-fitting model (and lowest DIC) was obtained for treatment model T3b with duration of treatment as a covariate. Note that, for this treatment comparison model, the study quality covariates did not yield a big improvement in model fit. This suggested that the choice of control treatment was confounded with other study quality covariates. In other words, this suggested that treatment effects were smaller when albumin was used as a control, indicating adequate blinding to treatment, plus also acting as a proxy for other aspects of study quality rather than albumin necessarily having a clinical effect on mortality as compared with no treatment.

Considering treatment comparison models T2 and T3b, the fixed-effects models giving the lowest DIC are highlighted in bold in Table 15 and are as follows:

• T3b with duration of treatment

• T2 with duration of treatment + Jadad score

• T2 with duration of treatment + publication date

• T2 with duration of treatment + 1/N

• T2 with daily dose + 1/N

• T2 with volume + 1/N.

There is little to choose between these models on the basis of model fit and DIC.

As described above, dropping dosing regime covariates from these models (owing to the lack of clinical interpretability) and replacing the heterogeneity explained by these covariates with a random-effects model with a between-study heterogeneity parameter was considered (below all random-effects models):

• T3b

• T2 with Jadad score

• T2 with publication date

• T2 with 1/N.

Fixed-effects model T3b with duration of treatment was selected as our best-fitting model, but results from the other nine models, listed above, as a sensitivity analysis to assess robustness of conclusions on clinical efficacy to model choice are also presented.

Best-fitting model – all-cause mortality

Our best-fitting treatment comparison model was model T3b with duration of treatment. Treatment comparison model T3b compares IVIG/IVIGAM versus albumin versus no treatment. As discussed previously, choice of control appeared to be a proxy for study quality. Therefore, the estimate of relative treatment effect used was for IVIG/IVIGAM versus albumin. Figure 29 shows the OR for IVIG/IVIGAM versus albumin, plotted against duration of treatment, with 95% credible intervals (figures for ORs are provided in Table 16). The treatment effect was stronger for longer durations of treatment; the majority of studies used a duration of treatment of 3 days. For a duration of treatment of 3 days, the OR of mortality for IVIG/IVIGAM versus albumin was 0.75 (95% credible interval 0.58 to 0.96), indicating that there was some evidence that IVIG/IVIGAM was effective in reducing all-cause mortality.

FIGURE 29.

Treatment comparison model T3b: IVIG/IVIGAM versus albumin with duration of treatment – OR (posterior mean and 95% credible intervals) by duration of treatment (days).

TABLE 16 - Predicted OR (95% credible intervals) for each of the 10 best-fitting treatment comparison models

Model	Covariate	Predicted OR (95% credible intervals)
Fixed-effects treatment comparison model T3b Covariate: duration of treatment	Duration (days):
	2	1.10 (0.79 to 1.44)
	3	0.75 (0.58 to 0.96)
	4	0.52 (0.37 to 0.72)
	5	0.37 (0.22 to 0.58)
	6	0.26 (0.13 to 0.47)
	7	0.19 (0.07 to 0.39)
Fixed-effects treatment comparison model T2 Covariates: duration of treatment and Jadad score	Duration (days):
	2	1.07 (0.79 to 1.43)
	3	0.81 (0.59 to 1.11)
	4	0.63 (0.39 to 0.95)
	5	0.49 (0.24 to 0.85)
	6	0.38 (0.15 to 0.77)
	7	0.30 (0.09 to 0.70)
Fixed effect Treatment comparison model T2 Covariates: duration of treatment and publication date	Duration (days):
	2	1.05 (0.78 to 1.38)
	3	0.82 (0.60 to 1.08)
	4	0.64 (0.41 to 0.96)
	5	0.51 (0.26 to 0.89)
	6	0.41 (0.17 to 0.84)
	7	0.33 (0.10 to 0.79)
Fixed-effects treatment comparison model T2 Covariates: duration of treatment and 1/N	N → ∞
	Duration (days):
	2	1.33 (0.84 to 2.01)
	3	1.07 (0.61 to 1.73)
	4	0.86 (0.41 to 1.58)
	5	0.71 (0.27 to 1.49)
	6	0.59 (0.17 to 1.43)
	7	0.50 (0.11 to 1.41)
	N = 339
	Duration (days):
	2	1.04 (0.76 to 1.39)
	3	0.83 (0.58 to 1.14)
	4	0.67 (0.39 to 1.05)
	5	0.54 (0.25 to 1.00)
	6	0.45 (0.16 to 0.98)
	7	0.38 (0.10 to 0.97)
Fixed-effects treatment comparison model T2 Covariates: daily dose and 1/N	Daily dose (g/kg/day):
	0.1	0.48 (0.22 to 0.89)
	0.2	0.60 (0.34 to 0.97)
	0.3	0.76 (0.52 to 1.07)
	0.4	1.00 (0.72 to 1.30)
	0.5	1.29 (0.87 to 1.83)
	0.6	1.70 (0.97 to 2.78)
	0.7	2.27 (1.04 to 4.38)
Fixed-effects treatment comparison model T2 Covariates: volume and 1/N	Volume (ml kg–1 day–1)
	2	0.48 (0.22 to 0.91)
	3	0.54 (0.28 to 0.94)
	4	0.61 (0.34 to 0.98)
	5	0.68 (0.43 to 1.02)
	6	0.77 (0.52 to 1.08)
	7	0.87 (0.63 to 1.17)
	8	0.99 (0.72 to 1.30)
	9	1.12 (0.81 to 1.52)
	10	1.28 (0.87 to 1.81)
Random-effects treatment comparison model T3b		0.68 (0.16 to 1.83)
Random-effects treatment comparison model T2 Covariate: Jadad score		0.83 (0.18 to 2.13)
Random-effects treatment comparison model T2 Covariate: publication date		0.83 (0.24 to 1.72)
Random-effects treatment comparison model T2 Covariate: 1/N	N → ∞	1.27 (0.25 to 3.17)
	N = 339	0.92 (0.23 to 2.10)

Sensitivity analyses for remaining nine treatment comparison models with covariates

In models with covariates, we need to choose a specific value of each covariate in order to obtain estimated treatment effects. First consider the fixed effect treatment comparison model T2 with covariates duration of treatment and Jadad score. Jadad score is an indicator for quality (Jadad score of 5 indicating best quality or lowest risk of bias).The Jadad score was fixed to 5 to produce the treatment estimates from this model, which gives a treatment effect estimate that can be interpreted as adjusting for bias introduced by study quality. Figure 30 presents the OR for IVIG/IVIGAM versus albumin/no treatment for 3 days, the OR of mortality for IVIG/IVIGAM versus albumin/no treatment was 0.81 (95% credible intervals 0.59 to 1.11), suggesting only weak evidence that IVIG/IVIGAM was effective in reducing all-cause mortality with the 95% credible intervals containing 1.00 (i.e. no effect).

FIGURE 30.

Treatment comparison model T2: IVIG/IVIGAM versus albumin/no treatment with duration of treatment and Jadad score = 5 – OR (posterior mean and 95% credible intervals) by duration of treatment (days).

For the fixed effect treatment comparison model T2 with covariates duration of treatment and publication date we need to specify a value for publication date. As publication date reflects changes in clinical practise over time, the treatment effect estimate from the most recent studies in our evidence should be used. Publication date was fixed to 2007 to produce the treatment estimates from this model. This can be interpreted as controlling for changes in clinical practice over time. Figure 31 presents the OR for IVIG/IVIGAM versus albumin/no treatment plotted against duration of treatment for publication date of 2007. For duration of treatment of 3 days, the OR of mortality for IVIG/IVIGAM versus albumin/no treatment was 0.82 (95% credible interval 0.60, 1.08), suggesting only weak evidence that IVIG/IVIGAM was effective in reducing all-cause mortality with the 95% credible intervals containing 1.00 (i.e. no effect).

FIGURE 31.

Treatment comparison model T2: IVIG/IVIGAM versus albumin/no treatment with duration of treatment and publication date of 2007 – OR (posterior mean and 95% credible intervals) by duration of treatment (days).

For the fixed-effects treatment comparison model T2 with covariates duration of treatment and 1/N, as sample size is an indicator for publication bias and other small-study effects,69,70 the treatment effect estimate from larger studies in/from our evidence should be used. Letting N → ∞ can be interpreted as representing the treatment effect estimated from an infinitely large study. This can be interpreted as adjusting for publication bias and other small-study effects. However, letting N → ∞ may lead to extrapolation well outside the limits of the data set with which the model was fitted. Results are also presented for N = 339, the largest treatment arm sample size in the set of studies included in our evidence synthesis. Figure 32a and b presents the ORs for IVIG/IVIGAM versus albumin/no treatment plotted against duration of treatment for N → ∞ and N = 339, respectively. For duration of treatment of 3 days and letting N → ∞, the OR of mortality for IVIG/IVIGAM versus albumin/no treatment was 1.07 (95% credible interval 0.61 to 1.73), showing no evidence that IVIG/IVIGAM was effective in reducing all-cause mortality. However, this estimate extrapolated the effects of sample size beyond the limits in the data set. For duration of treatment of 3 days and letting N = 339, the OR of mortality for IVIG/IVIGAM versus albumin/no treatment was 0.83 (95% credible interval 0.58 to 1.14), suggesting only weak evidence that IVIG/IVIGAM was effective in reducing all-cause mortality with the 95% credible interval containing 1.00 (i.e. no effect). Although N = 339 may appear an arbitrary choice, this should be interpreted as adjusting for publication bias/small-study effects based on the assumption that the largest trial published in this area was not subject to such bias. The results from using N = 339 are comparable to the other study quality adjustment results; therefore, this value was used for presenting the results of further models with 1/N as a covariate.

FIGURE 32.

Treatment comparison model T2: IVIG/IVIGAM versus albumin/no treatment with duration of treatment (a) for N ≥ ∞ (b) for N = 339 – OR (posterior mean and 95% credible intervals) by duration of treatment (days).

For the fixed-effects treatment comparison model T2 with covariates daily dose + 1/N, Figure 33 presents the OR for IVIG/IVIGAM versus albumin/no treatment plotted against daily dose of IVIG for N = 339. Treatment effect was stronger with lower daily doses. Average daily dose was approximately 0.3 g/kg/day. For a daily dose of 0.3 g/kg/day, the OR of mortality for IVIG/IVIGAM versus albumin/no treatment was 0.76 (95% credible interval 0.52 to 1.07), suggesting only weak evidence that IVIG/IVIGAM was effective in reducing all-cause mortality with the 95% credible intervals containing 1.00 (i.e. no effect).

FIGURE 33.

Treatment comparison model T2: IVIG/IVIGAM versus albumin/no treatment with daily dose for N = 339 – OR (posterior mean and 95% credible intervals) by daily dose (g/kg/day).

For the fixed-effects treatment comparison model T2 with covariates volume and 1/N, Figure 34 presents the OR for IVIG/IVIGAM versus albumin/no treatment plotted against volume of IVIG therapy for N = 339. Treatment effect was stronger with lower volumes. Average volume was approximately 5 ml/kg/day. For a volume of 5 ml/kg/day, the OR of mortality for IVIG/IVIGAM versus albumin/no treatment was 0.68 (95% credible interval 0.43 to 1.02), suggesting only weak evidence that IVIG/IVIGAM was effective in reducing all-cause mortality with the 95% credible interval containing 1.00 (i.e. no effect).

FIGURE 34.

Treatment comparison model T2: IVIG/IVIGAM versus albumin/no treatment with volume for N = 339 – OR (posterior mean and 95% credible intervals) by volume (ml/kg/day).

For the random-effects models it was assumed that the treatment effects for the different studies come from a common population of treatment effects. The predicted effect in a ‘new’ or ‘typical’ study, drawn from this common distribution, was used to summarise the treatment effect from these models. Note that this produced wider credible intervals than for the fixed-effects models because there were two components of uncertainty: one in the estimate of the pooled mean and the second in where the distribution of effects the population of interest might lie.

For the random-effects treatment comparison model T3b, the OR of mortality for IVIG/IVIGAM versus albumin was 0.68 (95% credible interval 0.16 to 1.83), suggesting a large degree of uncertainty that IVIG/IVIGAM was effective in reducing all-cause mortality.

For the random-effects treatment comparison model T2 with covariate Jadad score, for the Jadad score of 5, the OR of mortality for IVIG/IVIGAM versus albumin was 0.83 (95% credible interval 0.18 to 2.13), suggesting a large degree of uncertainty that IVIG/IVIGAM was effective in reducing all-cause mortality.

For the random-effects treatment comparison model T2 with covariate publication date, for a publication date of 2007, the OR of mortality for IVIG/IVIGAM versus albumin was 0.83 (95% credible interval 0.24 to 1.72), suggesting a large degree of uncertainty that IVIG/IVIGAM was effective in reducing all-cause mortality.

For the random-effects treatment comparison model T2 with covariate 1/N, for N → ∞, the OR of mortality for IVIG/IVIGAM versus albumin was 1.27 (95% credible interval 0.25 to 3.17) and for N = 339, the OR of mortality for IVIG/IVIGAM versus albumin was 0.92 (95% credible intervals 0.23 to 2.10). Both of these suggested a large degree of uncertainty that IVIG/IVIGAM was effective in reducing all-cause mortality. However, when N = 339, the results were in line with other results from the random-effects models, whereas when N → ∞, the posterior mean OR was > 1 and the 95% credible intervals were very wide.

Summary of sensitivity analyses

The treatment effect estimates were robust to the choice of method to adjust for study quality, publication bias and small-study effects (when N was set equal to the maximum arm size in the studies in our review – in those models that depended on sample size – avoiding extrapolation beyond the data set).

Treatment effect estimates were, however, sensitive to the assumed dose regime. It was not clear which values these should take. Robust results were obtained by setting these covariates equal to their average value. From the studies in our review, however, in the absence of any clinical rationale why these covariates should have a causative relationship with treatment efficacy, these relationships can only be considered as association. For this reason, the random-effects models that omitted these covariates were explored.

The results from the different random-effects models were fairly comparable, but provided wider credible intervals than the results from the fixed-effects models with the dosing regime covariates included. This was because the dosing covariates were not being used to explain heterogeneity but, instead, the heterogeneity present was acknowledged and a prediction made for a population drawn from the distribution of study effects.

Key findings

There is evidence that there are issues with bias associated with trial methodology and publication/small-study effects and these were, therefore, explored by adjusting treatment effect using measures of trial methodology or publication bias/small-study effects (Jadad score, publication date, sample size, choice of control). Results were found to be fairly robust to whichever measure of study quality was adjusted for (note, a marginally significant result can become a marginally non-significant result). The conclusion is that there is a borderline significant (at the 5% level) effect of IVIG on reducing all-cause mortality for patients with severe sepsis/septic shock.

There was a large degree of heterogeneity in the treatment effects between studies. However, some measure of dosing regime, together with a measure of study quality or study size, could explain the between-study heterogeneity in treatment effect results. The estimates of treatment effect were therefore sensitive to the dosing regime; however, there was no clear clinical rationale for this result.

The best-fitting model adjusted for study quality by incorporating an effect for the choice of control (albumin or no treatment) and included duration of IVIG therapy as a treatment effect modifying covariate. The resulting treatment effect estimates, therefore, depended on duration of IVIG therapy. The most commonly used duration of therapy reported in the studies was 3 days and so this was chosen to report the results. This gave an OR of 0.75 with a 95% credible interval of 0.58 to 0.96, showing a reduction in the odds of all-cause mortality in patients with severe sepsis/septic shock using IVIG compared with albumin, a result that was just marginally significant at the 5% significance level.

If the heterogeneity explained by duration of IVIG therapy was treated as unexplained heterogeneity (i.e. a random-effect models), the results still showed a reduction in the odds of all-cause mortality in patients with severe sepsis using IVIG compared with albumin (OR 0.68), but the 95% credible intervals were widened (0.16 to 1.83) such that this result was no longer statistically significant.

Comparison with previous meta-analyses

There have been several previous meta-analyses conducted on IVIG for severe sepsis/septic shock (see Tables 3 and 4) and conflicting conclusions have been drawn. 71 The different meta-analyses produce slightly different results owing to the included studies (see Table 4), the type of IVIG (or IVIGAM) included, whether and how ‘high-quality’ trials have been defined and how heterogeneity has been accounted for (whether with fixed- or random-effects models and whether or not treatment moderating covariates have been adjusted for).

Previous meta-analyses that have estimated treatment effects separately for IVIG and IVIGAM10,15 have found a strong treatment effect for IVIGAM and a borderline significant effect for IVIG. However, although Kreymann et al. 10 reported that this result was robust to including high-quality evidence only, Alejandria et al. 15 found that when they restricted their analysis to studies at low risk of bias only, then neither of the treatment effects for IVIG or IVIGAM were significant at the 5% level. Most of the previous meta-analyses that explored the effects of trial quality11,14,15 found significant treatment effects when all evidence was included, but non-significant results when the analyses were restricted to ‘high-quality’ trials, however defined.

Although all previous meta-analyses tested for heterogeneity, all (with the exception of Turgeon et al. 12) performed a fixed-effects meta-analysis. Turgeon et al. 12 fitted a random-effects model, to allow for the heterogeneity between studies, and also explored factors that may explain the heterogeneity in treatment effects between studies. They found that the following factors were associated with treatment effect: dosage regime, duration of IVIG therapy, trial quality, publication date and whether patients had septic shock or other forms of severe sepsis. These results are all in line with our findings. Laupland et al. 11 demonstrated that treatment effects were stronger when no treatment was used as the control compared with when albumin was used, again in line with our findings.

Our meta-analysis is the first to simultaneously allow for type of IVIG/IVIGAM, control treatment, study quality/publication bias, dosing regime and other potential covariates. When some measure of study quality (e.g. choice of control) and some measure of dosing regime (e.g. duration of IVIG therapy) were controlled for, there appeared to be no difference between the type of IVIG/IVIGAM therapy.

Limitations of available evidence

As has been identified in previous meta-analyses, there are issues with the methodological quality of the available evidence. Although the treatment effect results are fairly robust to various different approaches to adjust for these, because the treatment effect measure is ‘borderline significant at the 5% level’, the choice of method to adjust for study quality/publication bias can lead to either significant or non-significant results. Although we do not place too much focus on statistical significance and focus more on the credible intervals, this sensitivity to the method for adjusting for study quality/publication bias causes some concern with the interpretation of the results based on this evidence.

There is substantial heterogeneity in the treatment effects from the different studies. Although this heterogeneity can be explained using aspects of the dosing regime, on detailed discussion with the Expert Group, it is not clear if there is any clinical rationale for these effects. These results should therefore be interpreted with caution, and we should note that these effects are only associations and should not be interpreted as necessarily causative in the absence of well-designed, dose-ranging studies.

Recommendations for models to be used in sensitivity analyses for cost-effectiveness modelling

There is no clear one best-fitting model that makes clinical sense. Sensitivity analyses to model results were therefore recommended for the cost-effectiveness modelling. The sensitivity analyses performed in the clinical effectiveness work suggested that the method used to adjust for study quality was not important (as long as one approach was used). The exception to this was letting in models that adjusted for sample size. Either one of the dosing covariates should be included or a random-effects model fitted. Results were sensitive to this choice and this should be explored in sensitivity analyses.

Inclusion criteria

A broad range of studies were considered for inclusion in the assessment of cost-effectiveness, including economic evaluations conducted alongside trials, modelling studies and analyses of administrative databases. Only full economic evaluations that considered both costs and consequences (including cost-effectiveness, cost–utility and cost–benefit analysis) and compared two or more treatment options were included in the literature review.

Studies were identified by electronically searching the NHS Economic Evaluation Database (NHS EED) via the Cochrane Library (searched 2 October 2009). No date or language restrictions were used. The full search strategy is reported in Appendix 2. The reference lists from identified studies were also screened.

Two reviewers (PP and SJP) independently assessed all titles and abstracts. Full texts of titles/abstracts deemed relevant were retrieved and the full text was used for the final selection. Reasons for excluding the full-text studies were recorded.

All studies of IVIG for the treatment of severe sepsis were critically reviewed with the assistance of a quality assessment checklist for cost-effectiveness studies (see Appendix 3). 72 For those studies evaluating non-IVIG interventions, information was extracted on the comparators, study population, main analytical approaches (e.g. patient-level analysis/decision-analytic modelling), primary outcome specified for the economic analysis, details of adjustment for quality of life and costing approaches. This information is tabulated and summarised in the following sections.

The differences in approaches, assumptions and data sources were explored to inform the need for additional systematic reviews and/or further primary or secondary analyses. The findings from these reviews and analyses provided the basis for the development and population of the new decision model reported in Chapter 5.

Cost-effectiveness of intravenous immunoglobulin

The single published study of IVIG evaluated the cost-effectiveness of IVIGAM (Pentaglobin) as an adjunct to standard care, compared with standard care alone, in an adult critical-care unit population with severe sepsis or septic shock. 73 The analysis was undertaken from a German hospital perspective.

The study was based on a decision-analytic approach using a simple decision tree structure (Figure 36) to assess the costs and effects of IVIGAM added to standard care compared with standard care alone. Although the time horizon was not explicitly stated in the analysis, the model structure was restricted to the initial period in the critical-care unit only. Conditional upon the treatment strategy, the model estimates the associated probability of critical-care unit mortality/survival. These estimates were subsequently used to derive the number needed to treat (NNT) with IVIGAM to avoid one case of critical-care unit mortality being compared with standard care alone. Cost-effectiveness was assessed by estimating the incremental cost per additional life saved. No discounting of future costs or benefits was carried out given the focus on the initial period in the critical-care unit.

FIGURE 36.

Structure of the decision tree used for the evaluation of IVIGAM. ICU, intensive care unit.

Clinical evidence for the risk of critical-care unit mortality for IVIGAM and standard care alone was based on the results of a separate meta-analysis that updated a previously published analysis. 90 Studies were restricted to the nine RCTs90 in adult populations comparing IVIGAM with placebo (435 patients; control group, n = 212; IVIGAM, n = 223).

The overall critical-care unit mortality in the pooled control arms of the placebo studies was reported to be 44% (95% CI 33% to 52%). In the model this was assumed to represent the baseline risk of critical-care unit mortality for standard care alone. Owing to heterogeneity in the relative treatment effect reported across the individual RCTs, the authors performed a random-effects meta-analysis to estimate the relative effectiveness of IVIGAM in reducing critical-care unit mortality (relative risk 0.57; 95% CI 0.43 to 0.74). Applying the relative risk to the baseline mortality risk resulted in an absolute risk reduction of 19% (95% CI 14% to 22%) for IVIGAM compared with standard care alone. From this, the NNT with IVIGAM to save one life was estimated to be 5.2 (95% CI 4.0 to 9.0).

Only direct medical costs incurred in the critical-care unit were included in the evaluation. These included: length of critical-care unit stay, use of ‘block’ therapies (i.e. sepsis therapy, blood therapy, ventilation and renal therapy) and the acquisition costs of IVIGAM. The length of critical-care unit stay and unit costs were assumed to be different for survivors and non-survivors of the critical-care unit stay. The difference in the length of critical-care unit stay between IVIGAM and standard care was, therefore, based on the difference in the proportion of survivors and non-survivors between the two strategies. The acquisition costs of IVIGAM were assumed to be the same for survivors and non-survivors. As only short-term costs associated with the initial critical-care unit stay were considered, discounting was not applied. The price year for costs was not stated. The results of the base-case analysis showed that the use of IVIGAM resulted in incremental costs of €2037 compared with standard care alone.

The incremental cost-effectiveness ratio (ICER) for IVIGAM compared with standard therapy was €10,565 per life saved (i.e. €2037/0.1928). In a univariate sensitivity analysis (i.e. varying estimates for single input parameters), the ICER varied between €7231 and €28,443 per life saved. The results of the probabilistic sensitivity analysis (i.e. varying estimates for all input parameters simultaneously) suggested that at a willingness-to-pay level of €15,000 per life saved, the probability that IVIGAM is cost-effective was 83.9%.

Discussion and key issues

Only one published cost-effectiveness analysis of IVIG was identified,73 which evaluated the short-term cost-effectiveness of IVIGAM compared with standard care for severe sepsis/septic shock. The results of the quality assessment of this study are reported in Appendix 3. The quality assessment highlighted several important issues that potentially limit the generalisability of the findings from this study to UK clinical practice.

• The analysis has a short-term time horizon, restricted to the critical-care unit stay, and the long-term impact of IVIG has not been considered. Any assessment of the cost-effectiveness of IVIG should allow for the long-term cost and outcome implications of the short-term effects of the intervention. This ‘extrapolation’ is needed for two reasons. First, many patients who are treated for severe sepsis will continue to consume health-service resources beyond the initial critical-care unit stay and the cost-effectiveness of IVIG may influence these costs. Second, to compare the cost-effectiveness of IVIG with other uses of health-service resources (inside and outside of critical care), it is necessary to express the benefits of the drug in terms of a generic measure of health gain that can be compared across treatment areas. The most frequently used generic measure for this purpose is the quality-adjusted life-year (QALY). To provide a realistic estimate of the QALY impact of IVIG, the long-term implications for survival and health-related quality of life arising from the short-term effects of the intervention need to be incorporated.

• IVIGAM (Pentaglobin) is a specific IgM-enriched immunoglobulin product that is not available in the UK and the results of the cost-effectiveness analysis may not be generalisable to non-IgM-enriched immunoglobulin products. Consequently, it is not clear if IVIGAM is potentially more or less cost-effective than the alternative IVIG products available in the UK, owing to differences in clinical effectiveness and/or acquisition costs.

• Standard care has been used as comparator technology in the analysis. However, the authors do not attempt to define the components of standard care or to explore whether or not the placebo arms of the trials are likely to provide an appropriate source for this. As a result, it is difficult to assess the generalisability of the estimate of critical-care unit mortality to a UK setting.

• Both the baseline risk assigned to standard care and the relative effectiveness estimates of IVIGAM have been derived from the same studies included in the meta-analysis. Using the pooled estimate from the placebo arms of these studies to inform the baseline mortality risk for standard care arm, raises a couple of issues. First, as the studies included in the meta-analysis enrolled a small number of patients from various settings and had highly selected participants, it is unclear if the pooled baseline risk appropriately reflects standard care in Germany or the UK. Second, the heterogeneity noted across the studies in terms of the relative treatment effect may also exist for the baseline mortality risk in the placebo arms across the studies. Instead of using an ‘average’ measure of baseline risk, it might have been more appropriate to consider if some of the variation in mortality could be explained by study-level characteristics (e.g. severity of illness, setting) or if particular studies more closely related to the specific population and setting under consideration could have been used. Alternatively, external epidemiological evidence relevant to the setting and populations considered in routine clinical practice could have been more appropriate in informing the baseline risk of critical-care unit mortality (i.e. from cohort or registry data).

Cost-effectiveness of non-intravenous immunoglobulin interventions

The review of the cost-effectiveness evidence for IVIG identified several major limitations with the existing study and the results are unlikely to be relevant to informing the use of IVIG in the NHS for adult patients with severe sepsis and septic shock. Furthermore, the simple model structure, and the exclusion of long-term survival, quality of life and costs meant that this study did not provide much insight into the key areas required to develop our own model. Published decision-analytic models of other interventions for the management of severe sepsis were therefore examined.

The 15 studies74–89 identified evaluating the cost-effectiveness of other interventions for the management of adult patients with severe sepsis and septic shock are summarised in Table 17.

Twelve of these studies evaluated the cost-effectiveness of adding rhAPC to standard care compared with standard care alone. 74–86 Two studies evaluated the cost-effectiveness of early goal-directed therapy (EGDT)88,89 and one study evaluated the cost-effectiveness of systemic albumin infusion compared with standard medical care. 87

Two of the 12 rhAPC studies were from the UK. 80,82,83 The remainder were from the USA (n = 3),75–77 Canada (n = 2),74,86 France (n = 2),84,85 Germany (n = 1),78 Spain (n = 1)79 and Sweden (n = 1). 81 Both EGDT analyses were from the USA88,89 and the one albumin study was from France. 87 Eleven of the 15 studies reported results in terms of the incremental cost per QALY gained. 74,75,77,80–86,88,89

The perspective assumed by each of the studies is also reported in Table 17. The majority of studies considered the perspective of the providing institution (hospital) or of the health-care sector more generally. Four studies stated that a societal perspective was considered either in the base case or in a sensitivity analysis. 74,75,77,88 However, only one of these analyses74 actually attempted to incorporate the impact of productivity losses.

Model structures, time horizon and approaches to extrapolation

The studies used a range of different model structures and assumptions to model the costs and benefits of the interventions considered. Typically, studies used either (1) a simple decision tree; or (2) a combination of a decision tree to capture the short-term mortality of the initial episode and a Markov model to extrapolate survival and costs over a longer-term time horizon; or (3) a single Markov model to characterise both the short- and longer-term time horizons.

Markov models (Markov chains evaluated in discrete time) are useful when a decision problem involves modelling risk over time, when the timing of events is important and when these events may recur over time. 91 In contrast to a decision tree, where the full range of mutually exclusive pathways representing a patient’s prognosis are represented schematically by the individual tree ‘branches’, Markov models are based on a finite (or countable) number of discrete health states, called Markov states. In a Markov structure, hypothetical individuals reside in one out of the set of mutually exclusive health states at particular points in time. During discrete time intervals of equal length (normally referred to as Markov cycles), individuals can either remain in a particular health state or move to a separate health state (e.g. because of a patient experiencing a particular clinical event). The movements between states represent the potential clinical pathways that a patient may follow at different time points and over his or her remaining lifetime. The likelihood that an individual remains in a particular health state, or moves to a separate state, in the next Markov cycle is represented in terms of transition probabilities. Defining and subsequently estimating these transition probabilities represent both key structural and analytical elements of the decision model. The use of Markov model structure allows a more sophisticated approach to modelling the annual risk of death after the survival of the initial episode, allowing the annual risk of death to be altered over the longer term. Hence, the choice of model structure in the published studies relates closely to the study time horizon and the assumptions made to extrapolate short-term survival to mid- and long-term survival.

The majority of the studies used a decision tree approach to model short-term costs and outcomes of the alternative strategies. The short-term period varied in the studies between the critical-care unit stay, a fixed period of 28 days and/or the overall hospital stay. This period typically reflected the short-term nature of the relevant RCT evidence (and outcomes reported therein) used to inform the relative effectiveness of the specific interventions under investigation.

An example of a typical decision-tree structure is provided in Figure 37. Survival beyond the initial short-term period was estimated either by applying average age- and sex-specific estimates of the remaining years of life for short-term survivors or by adding a separate Markov model structure to the end of the short-term decision tree to characterise the longer-term prognosis of a sepsis survivor. Generally, a Markov model structure was used to estimate the duration over which a sepsis patient faced an increased risk of death compared with the general population (mid-term survival), after which the longer-term mortality risk was subsequently assumed to match that of the general population.

FIGURE 37.

Example of a short-term decision tree.

The majority of studies used the placebo arm of the Recombinant Human Activated Protein C Worldwide Evaluation in Severe Sepsis (PROWESS) study92 to estimate the baseline mortality risk in the short-term period assigned to standard care. Only two studies74,82,83 used epidemiological evidence from non-RCT sources to reflect the specific setting of the cost-effectiveness analysis. One study82,83 from the UK used national audit data from the ICNARC CMP Database and the authors of another study74 conducted a separate cohort study in critical-care units of three local tertiary care hospitals.

All but one76 of the studies considered time horizons beyond the short-term period and the majority of these used a lifetime horizon to estimate the remaining life-years and QALYs for a survivor of the short-term sepsis period. The remaining life expectancy for survivors of severe sepsis was calculated using two alternative approaches in the studies considered:

• the remaining life expectancy of survivors of severe sepsis was calculated in relation to the general population life expectancy by applying a single adjustment factor to represent the additional long-term mortality risk for sepsis survivors (e.g. relative survival over a lifetime was assumed to be approximately half that of the general population); or

• short-term survivors were assumed to have an increased risk of mortality for a specified time period ranging from 30 months to 8 years (mid-term survival), after which the mortality of survivors was generally assumed to match the mortality rates of the general population (long-term survival).

The estimates for the long-term increased risk of mortality for these approaches were typically taken from one of four studies reporting long-term survival rates for a cohort of sepsis patients or a general critical-care population. 74,93–95 The four studies and the methods used in the cost-effectiveness analysis to extrapolate the data are briefly described in Table 18.

Quality of life

Eleven66–75,77,80–85,88,89 of the 15 studies conducted cost–utility analysis and included an adjustment for quality of life of the remaining life expectancy estimates to estimate QALYs. Three80,81,89 of these 11 studies used utility values from a published abstract by Drabinski et al. 96 in adult sepsis survivors and four studies80,82–85 used values from a cohort study of ARDS by Angus et al. 97 Brief details of these two sources and the methods used by the cost-effectiveness studies are reported in Table 19. Typically, the remaining years of life were simply weighted by a single utility value (either 0.60 or 0.69) from one of these sources, to estimate lifetime QALYs for the survivors of severe sepsis.

In the remaining cost-effectiveness studies, two separate approaches were used to calculate the quality-adjusted survival of sepsis patients. Two cost-effectiveness analyses75,88 used the average quality-adjusted survival of someone in the general population with the same estimated remaining life expectancy and one study77 used published utility values for other non-sepsis health states, which were judged by the authors to be comparable to sepsis.

Resource use and costs

The key short-term and long-term costs included in the published cost-effectiveness studies are summarised in Table 20. All studies included the costs of study drug and the initial hospitalisation. Most of these studies separated the costs of the initial hospitalisation into critical-care unit and non-critical-care unit hospital ward costs. A limited number of studies also included the cost of treating adverse events, cost of other therapies (ventilation support, vasodilator support, renal support, blood therapy) and post-hospital costs up to day 28 (subsequent acute hospital care, nursing home, formal or informal supportive care at home).

Only 674,75,77,82,86,88 of the 15 studies modelled longer-term costs for survivors beyond the initial hospitalisation. Within these studies there was variation in both the types of costs included and the duration over which these costs were modelled. Although all of these studies included the cost of subsequent health care for survivors, fewer studies included other costs such as annual nursing home costs, productivity losses and death from any cause. Furthermore, only four75,77,82,83,88 of the six studies74–76,82,83,86,88 incorporated costs incurred over the remaining time horizon of the analysis. The two remaining studies restricted the analysis of long-term costs to a fixed period of 3 years. 74,86

Subgroup analysis

Ten74–79,81–84,86 of the 15 cost-effectiveness analysis also reported subgroup analysis. The most frequently reported subgroups included age, APACHE II score and number of organ dysfunctions.

Baseline mortality of severe sepsis/septic shock

Of the 16 studies73–89 considered in the cost-effectiveness review, the majority used data from the control arms of RCTs to estimate the baseline mortality risk during the critical-care unit or overall hospital stay. However, these RCTs were mainly or wholly undertaken outside the UK. In many respects, treatment patterns and resource use in the UK can be expected to differ from those in centres involved in the trials. One implication of these differences in UK practice is that the baseline event rates observed in the trials (i.e. in the control groups) are unlikely to provide reliable estimates for UK practice. For this reason, baseline mortality rates in our analysis were informed by additional primary data analysis of an alternative data source, the ICNARC Case Mix Programme (CMP) Database.

The CMP is the national comparative audit of patient outcomes from adult critical-care units in England, Wales and Northern Ireland, co-ordinated by ICNARC. The CMP is a voluntary performance assessment programme using high-quality clinical data to facilitate local quality improvement through routine feedback of comparative outcomes and key quality indicators to clinicians/managers in adult critical-care units. The CMP recruits predominantly adult general critical-care units, either standalone intensive care units or combined intensive care/high-dependency units. Currently, approximately 90% of adult, general critical-care units in England, Wales and Northern Ireland are participating in the CMP.

Case Mix Programme specified data are recorded prospectively and abstracted retrospectively by trained data collectors according to precise rules and definitions. Data collectors from each unit are trained prior to commencing data collection with retraining of existing staff, or training of new staff, also available. Data are collected on consecutive admissions to each participating critical-care unit and are submitted to ICNARC quarterly. Data are validated locally, on data entry, and then undergo extensive central validation for completeness, illogicalities and inconsistencies, with data validation reports returned to the units for correction and/or confirmation. The validation process is repeated until all queries have been resolved and then the data are incorporated into the CMP Database. The CMP Database has been evaluated according to the quality criteria of the Directory of Clinical Databases98 and scored highly. 99

Admissions with severe sepsis during the first 24 hours following admission to the critical-care unit were identified and extracted from the CMP Database using physiological criteria derived from those used in the PROWESS study of rhAPC. 3,5 Briefly, severe sepsis in the CMP is defined as evidence of infection (identified from the primary and/or secondary reason for admission to the critical-care unit), plus three or more SIRS criteria and at least one organ dysfunction (cardiovascular, respiratory, renal, haematological or metabolic) at any time during the 24-hour period. Severity of illness was summarised by the ICNARC physiology score,100 the APACHE II score26 and the number of organ dysfunctions. Outcome was measured by mortality at discharge from the original critical-care unit and mortality at ultimate discharge from an acute hospital. Activity was measured by length of stay in the critical-care unit (stratified by survival status at acute hospital discharge) and the total length of stay in acute hospital (stratified by survival status at acute hospital discharge). Both the outcome and activity data provide important sources relevant to informing parameter estimates for cost-effectiveness analysis in the UK. The database has been previously used to establish baseline epidemiology for severe sepsis in the UK3,5 and baseline event rates in one of the UK cost-effectiveness studies of rhAPC. 82,83

Data collected between 1995 and 2009 from the CMP Database were available for analysis. However, owing to changes in the management of patients in the UK, the mortality risk may have changed over time. This was explored descriptively by comparing critical-care unit and hospital mortality rates between 2002 and 2009 (Table 21). Given the trend towards lower mortality rates over time, analyses were restricted to data collected from 2007 to 2009.

Table 22 reports the mean sample characteristics used to inform the baseline mortality estimates in the decision model. The mean critical-care unit and overall hospital mortality for all admissions were 29.1% (95% CI 28.6% to 29.7%) and 40.6% (95% CI 40.0% to 41.2%), respectively. Variation in the baseline mortality risk in different subgroups may also have important implications in terms of cost-effectiveness assessments. A more detailed presentation of the baseline mortality across a range of subgroups is presented in Appendix 4. The use of these data within the cost-effectiveness model is discussed further in Chapter 5.

Long-term life expectancy

The assumptions used to estimate long-term survival of sepsis patients in the existing cost-effectiveness studies were typically based on three (non-UK) studies74,93,95 reporting long-term outcomes in severe sepsis patients and one UK study94 assessing long-term outcomes in general critical-care unit patients. As the assumptions and data sources were considered a key aspect of the development and population of our own cost-effectiveness model, an additional systematic search was conducted updating a previous review by Green et al. 82 In contrast to Green et al. ,82 cohort studies of general critical-care populations were excluded. These cohorts constitute a heterogeneous population and long-term mortality rates have been reported to vary significantly between patient subgroups. 101 Consequently, the assumption that average life expectancy of general critical-care unit patients reflects the life expectancy of a severe sepsis/septic shock patient admitted to a critical-care unit may not be appropriate. This review was, therefore, restricted to cohort studies of severe sepsis and septic shock patients.

Primary data analysis was also undertaken using an unpublished cohort study of severe sepsis from the UK. The results of the review and the primary data analysis are reported below.

Studies included in the systematic review

In addition to the two severe sepsis studies74,93 previously identified by Green et al. ,82,83 our updated MEDLINE search identified four new studies. 95,102–104 Two further studies were also identified in the hand search of relevant journals74,105 and three studies106–108 were identified from the reference lists of the included studies. The flow chart of studies included in the review is reported in Figure 38. Ten individual publications74,93,95,102,103,105–109 were identified from nine separate studies (with two publications coming from the same study).

FIGURE 38.

Flow chart of studies included in the long-term life expectancy review.

Additional primary data analysis

The cohort used was taken from a case–control study of the use of rhAPC in severe sepsis and septic shock. The data used were from the control group who did not receive rhAPC. This included 345 subjects from the Scottish Intensive Care Society (SICS) prospective, observational, multicentre, epidemiological study of sepsis in the Scottish critical-care database collected in a 5-month period in 2002. From these 345, only those patients (n = 271) for whom data on organ dysfunction were clearly reported were selected. The characteristics of this cohort (at admission to the critical-care unit) are shown in Table 25. Average follow-up for survival for this cohort was 787 (range 0–2062) days.

TABLE 25 - Characteristics of patients with severe sepsis, at admission, in the Cuthbertson data set (unpublished data) (with at least one recorded organ dysfunction)

OD, organ dysfunction.

Characteristic	Control (n = 271)
Age (years) on admission to the critical-care unit, mean (SD)	57.7 (14.1)
Male gender (%)	52.4
APACHE II score, mean (SD)	23.1 (8.4)
Quartiles of APACHE II score
First quartile	0–20
Second quartile	21–24
Third quartile	25–28
Fourth quartile	29–48
APACHE II score ≥ 25 (%)	47.6
Organ dysfunctions
Metabolic acidosis (%)	46.5
Haematological (%)	24.7
Renal (%)	41.3
Respiratory (%)	78.2
Cardiovascular (%)	56.8
Cardiovascular and renal OD (%)	26.9
Number of organ dysfunctions
One (%)	28.0
Two (%)	28.4
Three (%)	18.8
Four (%)	17.3
Five (%)	7.4
Length of stay (days) in the critical-care unit, mean (SD)	11.4 (12.9)
Mortality
Critical-care unit (%)	41.7
Original hospital (%)	47.6
Any hospital (%)	50.0
Overall follow-up period (%)	65.7
Follow-up (days) , mean (min–max)	787 (0–2062)

After hospital discharge, 144 subjects were alive and followed up. Figure 39 reports the Kaplan–Meier curve (and 95% CI) for survival after hospital discharge. The analytical approach used to populate the cost-effectiveness model with these data is discussed in Chapter 5.

FIGURE 39.

Observed survival after discharge from hospital (Kaplan–Meier).

Health-related quality of life after survival of severe sepsis (utilities)

Green et al. 82 have previously reported the findings from a literature search of health-state utilities associated with severe sepsis. However, this review identified only a single published abstract relating to patients with severe sepsis and septic shock. In the absence of robust data from the previous review, a separate systematic review was undertaken to update these findings.

Results

Four studies (including one abstract) were identified in our update review. 96,109–111 The flow chart of the search results and inclusion/exclusion of the studies is provided in Figure 40. A summary of these studies is reported in Table 26.

FIGURE 40.

Flow chart of studies included in the literature review for health-related quality of life (utilities). SF-36, Short Form questionnaire-36 items.

TABLE 26 - Characteristics of studies reporting utility data

NR, not reported.

n = 93.

n = 104.

Median.

n = 66.

n = 252.

Study	Country	Setting	n	Assessment times (number of respondents)	Male (%)	Mean age (years)	Mean APACHE II score
Drabinski et al. (2001)96	USA	53 hospitals	703	Day 30 (93)	52a	60a	NR
				Day 60 (93)
				Day 90 (93)
				Day 180 (93)
Granja et al. (2004)110	Portugal	1 critical-care unit	305	6 months (104)	64b	52b,c	17b,c
Korošec Jagodic et al. (2006)111	Slovenia	1 critical-care unit	66	2 years (10)	49d	64.4d	15.5d
Karlsson et al. (2009)109	Finland	24 critical-care units	470	Before critical illness (252)	64.7e	60.4c,e	24c,e
				≈ 17 months (156)

In all four studies,96,109–111 utilities were derived using the EQ-5D instrument. The EQ-5D is a standardised instrument for use as a measure of health outcome and is applicable to a wide range of health conditions and treatments. 112 It provides a simple descriptive profile and a single index value (utility) for health status. It is the measure currently recommended by the National Institute for Health and Clinical Excellence (NICE) to be used as part of its ‘reference case’ approach to undertaking cost-effectiveness analyses. 113

The follow-up period in the studies ranged from 6 months to 2 years. None of the studies reported data beyond 2 years’ follow-up. The utility values are reported in Table 27. Anchor points for these values are perfect health (1) and death (0). Differences in the patient characteristics, country, centres and assessment times across the studies again make a direct comparison problematic.

TABLE 27 - Comparison of utility values reported at follow-up points

Study	Follow-up point
	Before critical illness	1 month	2 months	3 months	6 months	17 months	2 years
Drabinski et al. (2001)96		0.53	0.62	0.68	0.69
Granja et al. (2004)110					0.84
Korošec Jagodic et al. (2006)111							0.72
Karlsson et al. (2009)109	0.70					0.86

Three109–111 of the four96,109–111 studies compared the utility values of sepsis patients with those of non-septic populations. Two110,111 of these studies compared the utility values of sepsis survivors with those of another critical-care unit population (patients admitted without sepsis110 or trauma patients111). Only one study109 compared the utility values of severe sepsis patients with age- and sex-adjusted general population estimates. In this study, the utility values of severe sepsis patients were reported to be lower than those of the age- and sex-adjusted general population, both before the onset of the clinical illness and at approximately 1.5 years after discharge from intensive care.

Two studies96,109 reported patients’ utility values for multiple time points, informing how the quality of life of sepsis survivors might alter over time. The abstract by Drabinski et al. 96 (previously identified by Green et al. 82) assessed changes in health status at days 30, 60, 90 and 180, suggesting that quality of life appears to improve within the first few months after a sepsis episode and then appears to plateau between 3 months and 6 months after the episode. Karlsson et al. 109 assessed patients’ health status within 1 week from the study entry and approximately 1.5 years after the study entry. Again, improvements in the quality of life were reported over the follow-up period. However, the majority (61.9%) of the initial assessments were completed by the patients’ next of kin to assess the patients’ quality of life before the sepsis episode. Consequently, the study does not directly inform the impact of the initial episode, nor does it provide an appropriate basis for assessing how quality of life may alter over the longer term.

Cost-effectiveness analysis

The decision model was developed and populated using data identified during phase I and the results from the clinical effectiveness review. All stages of the work were also informed by discussions with the Expert Group to provide feedback on specific aspects of the analysis including the model structure, data inputs and assumptions.

The model evaluated costs from the perspective of the NHS and Personal Social Services, expressed in UK pounds sterling at a 2009 price base. Outcomes were expressed in terms of QALYs. Both costs and outcomes were discounted using a 3.5% annual discount rate, in line with current guidelines. 113

The model was developed in the statistical programming package R114 and is probabilistic in that input parameters are entered into the model as probability distributions to reflect parameter uncertainty, i.e. uncertainty in the expected value of the inputs. 115 Monte Carlo simulation was used (5000 iterations) to propagate uncertainty in input parameters through the model in such a way that the results of the analysis can also be presented with their associated uncertainty. The probabilistic analysis also provided a formal approach to quantifying the consequences associated with the uncertainty surrounding the model results and can be used to identify priorities for future research. 116

The expected cost and QALYs for each of the strategies were estimated and compared, using ICERs where appropriate. The ICER represents the incremental cost per additional QALY associated with a more costly and effective strategy. The ICER was then compared against thresholds used by NICE to establish value for money in the NHS (currently in the region of £20,000–£30,000 per additional QALY). 113 These thresholds can be used to identify the optimal strategy in terms of cost-effectiveness considerations based on existing evidence.

A range of separate scenarios were also undertaken to assess the impact of key uncertainties related to input parameters and assumptions. Consistent with available evidence, the model also explored variability in the cost-effectiveness estimates for specific subgroups of patients.

Value of information analysis

Formal methods, based on EVI approaches, were used to identify potential research priorities and to establish whether or not investment in a multicentre RCT is likely to be a cost-effective use of resources. 116–118 The EVI approaches were also extended to consider a range of sources of uncertainty in the model to help identify and prioritise specific research questions that could also be addressed with other (non-RCT) research designs. The methods and results of this analysis are reported in Chapter 6.

The following sections outline the decision problem and the structure of the model, and report the key assumptions and data used to populate the model.

Treatment strategies/comparators

The decision problem addressed by the model relates to the cost-effectiveness of IVIG as an adjunctive treatment to standard care for the management of adults with severe sepsis and septic shock, compared with standard care alone. The base-case population in the model reflects the baseline characteristics of the population in the ICNARC CMP Database, under the assumption that this population is more representative of current NHS practice than the populations recruited into the RCTs. The impact of patient heterogeneity (e.g. owing to different clinical characteristics) was explored in separate analyses. This approach ensures that uncertainty in the decision because of the imprecision in parameter inputs can be separated from uncertainty in whether or not an intervention is cost-effective for particular subgroups of the population.

Model structure

The model evaluated the lifetime prognosis of severe sepsis in order to capture the long-term costs and consequences associated with the natural history of these patients in the absence of IVIG. The findings from the clinical effectiveness review were then employed to model the effect of using IVIG as an adjunctive treatment to standard care. The model structure was informed by the series of reviews described in previous chapters and is used to estimate lifetime costs and benefits associated with the primary outcome of the clinical effectiveness review: short-term all-cause mortality.

A simplified schematic of the decision model structure is shown in Figure 41 and a full technical description is provided in Appendix 5.

FIGURE 41.

Structure of the decision model representing the progression of patients diagnosed with severe sepsis or septic shock.

In common with many of the existing model structures, two related elements were considered reflecting short- and long-term consequences.

1. Short term The short-term consequences of the initial sepsis episode reflect the initial hospitalisation period (critical-care unit and non-critical-care unit). The decision tree quantifies the probability of surviving or dying during the initial hospitalisation for the sepsis episode. Baseline mortality data from the ICNARC CMP Database were used to estimate the risk of mortality associated with standard care and the results of the clinical effectiveness review were applied to estimate the risk with IVIG.

2. Long term Conditional on having survived the initial hospitalisation, a Markov structure was used to characterise the long-term prognosis over the remainder of a patient’s lifetime. Annual cycles were employed to reflect the annual probability of death for each year after the initial episode. Hence, the extent to which the use of IVIG reduces the risk of mortality during the initial hospitalisation period is translated into differences in long-term costs and QALYs on the basis of the long-term model.

In developing and populating the decision model there were two important considerations applied to inform the approaches and methods employed:

• the requirement to extrapolate outcomes beyond the time horizon of the main RCTs to ensure that differences in QALYs were appropriately quantified

• the need to ensure that the data inputs and assumptions were relevant to the specific population and setting to inform decision-making in the context of the NHS.

The use of decision analysis provides a number of advantages in exploring these issues in more detail: (1) it provides a framework to model both the short- and long-term costs and benefits associated with IVIG; (2) it makes each of these assumptions explicit and can highlight where the current uncertainties exist; (3) it provides a quantitative approach to combining evidence from separate sources and the use of probabilistic analysis means that the degree of uncertainty surrounding particular inputs can be reflected; (4) the potential impact of the assumptions on the cost-effectiveness of IVIG can be considered; and (5) the value of additional research to inform the decision problem can be established.

The following sections provide an overview of the model inputs and the methods used to inform the cost-effectiveness of IVIG. This information is summarised in Appendix 9 for the overall severe sepsis and septic shock population.

Model inputs

Long-term survival for sepsis survivors

UK data were used to estimate long-term survival for sepsis survivors from the cohort of the SICS prospective, observational, multicentre, epidemiological study of sepsis in Scottish critical care (Brian Cuthbertson, Sunnybrook Health Sciences Centre, 2010, personal communication) reported in Chapter 4. Parametric survival analyses were undertaken to estimate the long-term mortality estimates applied in the model using alternative distributions (Weibull, exponential and log-normal). To assess goodness of fit, the Akaike Information Criterion (AIC) was utilised along with a graphical inspection of the fit of the data and plausibility of longer-term predictions beyond the 5-year follow up period of the cohort study, before selecting the most appropriate curve for the final model.

Three separate models were fitted including additional covariates for:

1. age at admission

2. APACHE II score at admission, and

3. organ dysfunction (and age at admission).

The covariates were included to consider whether or not subgroup-specific estimates for the long-term survival were appropriate and to adjust for any potential imbalance between the baseline characteristics of the CMP data (used to estimate short-term mortality data) and the SICS cohort. Additional covariates considered were explored, but only age and APACHE II score were identified as significant predictors of long-term mortality (p < 0.05).

The full results from the parametric survival analysis are reported in Appendix 7. The distribution with the lowest AIC and representing the best statistical goodness of fit (sustained across different covariate sets) was the Weibull function. A graphical comparison of the predicted survival from the different parametric functions compared with the observed Kaplan–Meier survival curve (with 95% confidence bounds) is presented in Figure 42.

FIGURE 42.

Comparison of predicted survival from alternative parametric survival functions and observed Kaplan–Meier survival curve (with 95% confidence bounds).

The plausibility of the different parametric predictions beyond the 5 years of observed data was also explored by comparing these with age-adjusted estimates from the general population. This comparison is shown in Figure 43. It was considered implausible that the long-term mortality estimate of sepsis patients would become lower than that of the general population. Consequently, in the model, it was further assumed that the probability of mortality would be the maximum of that predicted from the parametric distributions and the observed yearly probability of mortality for the general population (age and sex adjusted). The time point at which the model switches from predictions from the parametric distributions to the estimates from the general population represents the point at which the mortality of the sepsis cohort is assumed to be the same as the mortality of the general population. For the overall population, the switch points were 9, 13 and 22 years for the log-normal, Weibull and exponential distributions, respectively. The ‘modified’ parametric survival functions are reported in Figure 44.

FIGURE 43.

Comparison of parametric survival functions over time with general population (GP) estimates.

FIGURE 44.

Comparison of ‘modified’ parametric survival functions over time with general population (GP) estimates.

Given the inevitable uncertainty about the longer-term survival extrapolation, the robustness of the results was explored using a range of scenario analyses in which we varied the time point at which patients switched from the predicted survival distributions to the corresponding estimates from the general population (varied between 5 years and 25 years).

Resource use and unit costs

Resource use and costs were estimated both for the short-term hospitalisation period and for the longer-term extrapolation. Costs assigned in the short-term period of the model included the acquisition costs of IVIG treatment and length of stay in hospital (critical-care unit and other wards). Costs assigned in the longer-term extrapolation were based on an assessment of the continuing costs of managing survivors after the initial hospitalisation.

The acquisition costs of IVIG were estimated from the cost per gram of products (5% concentration) reported in the British National Formulary (BNF). 120 The products and average costs are reported in Table 30. The total number of grams used was based on a 2 g/g dose assuming a weight of 70 kg and a duration of 3 days, based on advice from the Expert Group. This was rounded down to the nearest whole vial based on the current guidelines for use. 17 The total acquisition cost of IVIG was estimated to be £5539.

The length of stay in hospital (critical-care unit/non-critical-care unit) was informed using activity data from the same CMP Database used to estimate baseline mortality. Estimates of the mean and SE length of stay for survivors and non-survivors of the initial hospitalisation were used to inform the model input parameters. The length of stay in non-critical care wards was assumed to be the difference between the length of the overall hospitalisation and the length of critical-care unit stay. Table 31 reports the descriptive statistics based on all admissions. Results for key subgroups are reported separately in Appendix 8.

A per diem cost of £1293 was applied to the duration of the critical-care unit stay based on national unit cost estimates [National Schedule of Reference Costs 2007/08: NHS Trusts and Primary Care Trusts (PCTs) combined – Critical Care Services – Adult: Intensive Therapy Unit]. 121 For the remaining non-critical-care unit stay, a per diem estimate of £196 was used based on a general ward stay for septicaemia (National Schedule of Reference Costs 2007/08: NHS Trusts and PCTs combined Non-Elective Inpatient – Long Stay Excess Bed Day). 121

Existing cost-effectiveness studies were reviewed and additional searches were undertaken to identify potential sources of long-term cost data for the management of survivors of sepsis. Only one study74 was identified that reported estimates of the long-term costs for survivors of sepsis after the initial hospitalisation. This was a Canadian cohort study that reported costs in the first 3 years following the initial hospitalisation. Costs in the first year were reported to be considerably higher than those reported in years 2 and 3. Estimates reported in years 2 and 3 were very similar, suggesting that resource utilisation over the longer term was more stable. In the absence of any equivalent UK estimates, these estimates were converted to UK pounds sterling (and uprated to current prices). Given the similarity in the costs reported for years 2 and 3, these estimates were averaged and applied as an average cost incurred yearly in year 2 and beyond in the model. The specific estimates applied were £13,654 in the first year after discharge and £4467 for each year thereafter. In the absence of any reported measure of uncertainty around this estimate, a coefficient of variation of 2 was assumed. By using truncation, it was assumed that uncertainty could not lead to consider costs lower than the average annual per capita NHS cost of £1807.84.

Given the lack of UK cost data on the long-term management of sepsis survivors, additional scenarios were undertaken to explore the robustness of the model to alternative assumptions. For these scenarios alternative assumptions were explored regarding the magnitude of these estimates (± 50%). Given the lack of long-term UK cost data, the impact of alternative approaches using general population estimates of the average annual per capita NHS cost instead of using sepsis specific estimates was also explored.

Alternative clinical effectiveness scenarios

Table 32 reports the cost-effectiveness results using the best-fitting clinical effectiveness model for all-cause mortality [fixed-effects model T3b with covariate: duration of IVIG (3 days)]. The results show that the ICER of IVIG is £20,850 per QALY (i.e. incremental costs = £9308/incremental QALYs = 0.45), which is within the borderline region of estimates considered to be cost-effective in the NHS. At a threshold of £20,000 per QALY, the probability that IVIG is more cost-effective than standard care alone is 0.505. As the threshold cost per QALY increases, the probability that IVIG is cost-effective increases (i.e. increasing to 0.789 at a threshold of £30,000). The relationship between the threshold ICER and the probability that IVIG is cost-effective is shown more clearly in the CEAC reported in Figure 45.

FIGURE 45.

Cost-effectiveness acceptability curve using best-fitting model from the clinical effectiveness review.

Table 33 reports the results using each of the alternative models from the clinical effectiveness review considered within the sensitivity analysis. The ICER estimates vary between £16,177 per QALY to IVIG being dominated by standard care alone (i.e. IVIG being both less effective and more costly).

These results clearly demonstrate that any conclusions regarding the cost-effectiveness of IVIG are highly sensitive to the choice of model used for clinical effectiveness. The most favourable ICER estimate (£16,177) is obtained using a random-effects model (comparing IVIG with albumin). However, IVIG appears dominated when a random-effects model is used with an adjustment for publication bias using sample size (N) and setting N to infinity. As noted in the clinical review, setting N to infinity involves extrapolation beyond the data set; when N was restricted to 339 (i.e. equivalent to the largest existing study), then the ICER of IVIG was £28,520 per QALY.

Alternative long-term survival scenarios

Given the uncertainty surrounding the long-term survival extrapolation required to estimate lifetime QALY gains, the robustness of the results to alternative assumptions was explored. Two separate scenarios were considered.

1. Alternative time horizons Our main analysis was based on a lifetime time horizon (30 years) requiring extrapolation beyond the 5-year follow-up from our cohort study. The impact of restricting the analysis to shorter time horizons of between 5 years and 30 years was explored. This provides an indication of the importance of the period of extrapolation beyond the observed data in determining the overall cost-effectiveness of IVIG.

2. Long-term survival of sepsis patients compared with the general population The time point at which we assumed patients revert from the predicted survival distributions from the long-term cohort data to survival estimates from the general population was varied. In our main analysis, this time point was determined by the time the mortality predictions from the parametric survival analysis became lower than the equivalent age- and sex-matched estimates from the general population. In the separate scenarios, this switch was assumed to happen at fixed time points between 5 years and 25 years after the initial episode.

Both scenarios were undertaken using the estimate of short-term clinical effectiveness from the fixed-effect model T3b with covariate: duration of IVIG (3 days).

Table 34 reports the cost-effectiveness results based on alternative time horizons. Restricting the time horizon to 5 years increased the ICER of IVIG to £43,717 per additional QALY, well above the conventional threshold considered to represent value for money to the NHS. As the time horizon increased, the cost-effectiveness of IVIG became more favourable. The results clearly demonstrate that the cost-effectiveness of IVIG is dependent upon the additional QALY gains predicted as part of the longer-term extrapolation.

Table 35 reports the cost-effectiveness results based on varying the time point at which sepsis survivors are assumed to revert back to general population mortality rates. The ICER estimates improved marginally when it was assumed that patients reverted back to general population mortality rates earlier than in our main analysis. When it was assumed that patients reverted back to the general population mortality rate immediately after the 5-year follow-up period of the separate cohort study, the ICER improved to £19,974 per QALY, just under the lower bound of current cost-effectiveness thresholds. However, the ICER estimate increased to over £20,000 for all other time points. The ICER varied between £19,974 and £20,164 across these scenarios, indicating that the assumption that the prognosis of severe sepsis patients remains worse than that of the general population after 5 years is not a key driver of cost-effectiveness.

Alternative long-term cost scenarios

Given the lack of UK cost data on the long-term management of sepsis survivors, additional scenarios were undertaken to explore the robustness of the model to alternative costing assumptions. The impact of the following approaches was considered: (1) altering the magnitude of these estimates (± 50%) and (2) using general population estimates of the average annual per capital NHS cost, instead of sepsis-specific estimates from a non-UK source. Again, both scenarios were undertaken using the estimate of short-term clinical effectiveness from the fixed-effects model T3b with covariate duration of IVIG (3 days).

Table 36 reports the cost-effectiveness results varied the estimates applied in the main analysis by ± 50%. The ICER across this range varied between £19,418 and £22,282 per QALY, suggesting that the results were relatively robust to this magnitude of change. However, the ICER estimates appeared more sensitive to the assumption that the long-term costs of managing sepsis survivors would be higher than the annual NHS costs incurred by the general population in the longer term. Table 36 presents the ICER results assuming that the long-term costs of sepsis patients were (1) the same the general population after 3 years (reflecting the time horizon of the study used to estimate the long-term costs) or (2) the same as the general population immediately after the initial hospitalisation period. The ICERs for these separate analyses were £17,962 and £15,792 per QALY, respectively. Both of these estimates were well within the threshold range considered to represent value for money to the NHS, suggesting that the assumption that patients continue to incur higher costs than the general population over the longer-term extrapolation period is an important consideration.

Subgroups

The results of the scenarios presented have been based on the average baseline characteristics of patients in the ICNARC CMP Database. However, the cost-effectiveness results may also vary according to different patient characteristics. Heterogeneity in patient characteristics and the impact on the ICER estimates were explored using a series of separate scenarios based on:

• APACHE II score

• ICNARC physiology score

• organ dysfunctions.

These scenarios were explored by varying the baseline hospital mortality rate according to the particular characteristics considered (using predictions from the logistic regressions detailed in Chapter 4 and Appendix 6). Long-term mortality was also varied according to APACHE II score as this score was demonstrated to be a significant predictor of long-term mortality in our long-term cohort data (using predictions from the parametric survival regressions detailed in Chapter 5 and Appendix 6). Subgroup estimates of long-term mortality were not used for either the analyses based on ICNARC physiology score, as these data were not collected within the long-term cohort, or those based on organ dysfunctions, as these covariates were not identified as significant predictors of long-term mortality. All scenarios were undertaken using the estimate of short-term clinical effectiveness from the fixed-effects model T3b with covariate duration of IVIG (3 days).

Detailed results based on the APACHE II and ICNARC physiology scores are presented in Appendix 10. A clear non-linear relationship is apparent in the cost-effectiveness estimates. That is, ICER estimates were markedly higher for very low- and very high-risk patients compared with our main results. However, these higher estimates were reported for relatively extreme scores, which were not considered representative of the majority of patients in the CMP Database. The results indicated that the cost-effectiveness results were relatively robust to variation in the scores actually observed in the database.

Table 37 presents the cost-effectiveness results for different subgroups defined according to the number of organ dysfunctions during the first 24 hours following admission to the critical-care unit. The ICER estimates for IVIG were more favourable for patients with two or more organ dysfunctions (£20,706) than for those with only one (£26,049). However, simply dichotomising the population in this manner ignored the potential heterogeneity that existed within the group with two or more organ dysfunctions. Within this subgroup, the ICER varied between £20,611 (three organ dysfunctions) and £26,268 (five organ dysfunctions).

Recommendation 1

Research on the mechanism(s) of action of IVIG preparation(s) in the severe sepsis population commencing with a thorough review of existing research prior to embarking on any new research.

Recommendation 2

Informed by recommendation 1, dose-ranging/finding studies to identify dose, timing of dose and safety data (tolerability/side-effects) to inform the intervention(s) for a future multicentre RCT.

There was a dearth of long-term outcome and cost/resource data on severe sepsis survivors to inform the cost-effectiveness analyses. Either by exploiting existing databases, through record linkage, or by initiating a prospective cohort study, long-term survival, including quality of survival and costs of survival for several years after the initial severe sepsis episode, should be explored.

Recommendation 3

Research to inform the long-term survival, including quality and costs of survival for the severe sepsis population.

Recommendation 4

Results of recommendations 1–3 should be re-evaluated for their impact on our EVI analyses.

The primary target population is adult patients with severe sepsis. There is increasing awareness that the syndrome described as severe sepsis represents a large and extremely heterogeneous group of patients. The heterogeneity of the severe sepsis population has plagued large, multicentre RCTs and there is a realisation that the focus should be on more homogeneous, specific, severe sepsis subpopulations. Heterogeneity appears to exist at the genetic, biochemical and clinical level, all of which may be associated. The current focus of research on severe sepsis has been in the identification of relevant genetic, biochemical and clinical markers with the aim better describing more homogeneous severe sepsis subpopulations, providing more rapid bedside markers for the early identification of sepsis and establishing which patients are most likely to benefit from therapy. Such advancements should inform the final design of a multicentre RCT of IVIG.

Recommendation 5

Recommendations 1–3 require knowledge of, and design of the definitive RCT for IVIG in severe sepsis requires a comprehensive review of, the emerging evidence surrounding the heterogeneity of the severe sepsis population at the genetic, biochemical and clinical level.

In summary, although the results highlight the value for money obtained in conducting further primary research in this area, the biggest limitation for such research regards the uncertainties over the heterogeneous nature of severe sepsis and the mechanism of action of IVIG. Resolving these would allow for better definition of the plausibility of the effectiveness scenarios presented and, consequently, a better understanding of the cost-effectiveness of this treatment. This information would also inform the design of future, primary evaluative research.

MEDLINE

Database: Ovid MEDLINE(R) In-Process & Other Non-Indexed Citations and Ovid MEDLINE(R) < 1950 to present >

Searched via Ovid interface: 2 October 2009

1. immunoglobulins/ (37,225)

2. immunoglobulin$.tw. (105,959)

3. ivig.tw. (3259)

4. 1 or 2 or 3 (125,909)

5. sepsis/ (34,958)

6. sepsis.tw. (48,066)

7. septic shock/ (15,622)

8. septic shock.tw. (10,633)

9. septicemia/ (34,958)

10. septicaemia.tw. (4736)

11. septicemia.tw. (9296)

12. 5 or 6 or 7 or 8 or 9 or 10 or 11 (91,020)

13. 4 and 12 (1290)

14. randomized controlled trial.pt. (283,692)

15. controlled clinical trial.pt. (80,983)

16. randomized.ab. (201,142)

17. placebo.ab. (120,675)

18. drug therapy.fs. (1,358,908)

19. randomly.ab. (148,294)

20. trial.ab. (208,511)

21. groups.ab. (997,172)

22. 14 or 15 or 16 or 17 or 18 or 19 or 20 or 21 (2,559,489)

23. exp animals/not humans.sh. (3,473,795)

24. 22 not 23 (2,179,816)

25. 13 and 24 (388)

26. limit 25 to yr=“2001 -Current” (160)

EMBASE

Database: EMBASE < 1996 to 2009 week 39 >

Searched via Ovid interface: 2 October 2009

1. immunoglobulins/ (28,604)

2. immunoglobulin$.tw. (40,750)

3. ivig.tw. (2552)

4. 1 or 2 or 3 (58,658)

5. sepsis/ (31,978)

6. sepsis.tw. (26,319)

7. septic shock/ (10,171)

8. septic shock.tw. (6530)

9. septicemia/ (5658)

10. septicaemia.tw. (1560)

11. septicemia.tw. (2717)

12. 5 or 6 or 7 or 8 or 9 or 10 or 11 (53,621)

13. 4 and 12 (1367)

14. random.tw. (61,491)

15. placebo.mp. (123,372)

16. double-blind.tw. (47,097)

17. 14 or 15 or 16 (196,237)

18. 17 and 13 (111)

19. limit 18 to yr=“2001 -Current” (94)

20. animals/not (animals/and humans/) (2161)

21. 19 not 20 (95)

NHS Economic Evaluation Database

Searched via The Cochrane Library (www.mrw.interscience.wiley.com/cochrane/cochrane_search_fs.html): 2 October 2009

#1 MeSH descriptor Immunoglobulins, this term only

#2 (immunoglobulin*)

#3 (ivig)

#4 (#1 OR #2 OR #3)

#5 MeSH descriptor Sepsis, this term only

#6 (sepsis)

#7 MeSH descriptor Shock, Septic, this term only

#8 (septic shock)

#9 MeSH descriptor Hemorrhagic Septicemia, this term only

#10 (septicaemia)

#11 (septicemia)

#12 (#5 OR #6 OR #7 OR #8 OR #9 OR #10 OR #11)

#13 (#4 AND #12)

Results for all sepsis (line #12) and for sepsis and IVIG (line #13) saved.

MEDLINE

Database: Ovid MEDLINE(R) In-Process & Other Non-Indexed Citations and Ovid MEDLINE(R) < 1950 to present >

Searched via Ovid interface: 20 October 2009

1. sepsis/ (35,031)

2. sepsis.tw. (48,315)

3. shock, septic/ (15,639)

4. septic shock.tw. (10,675)

5. septicaemia.tw. (4752)

6. septicemia.tw. (9308)

7. 1 or 2 or 3 or 4 or 5 or 6 (91,334)

8. prognos$.tw. (269,223)

9. first episode.tw. (5305)

10. cohort.tw. (141,155)

11. 8 or 9 or 10 (405,343)

12. 7 and 11 (5309)

13. limit 12 to yr=“2004 -Current” (2154)

14. exp animals/not humans/ (3,478,640)

15. 13 not 14 (2097)

MEDLINE

Database: Ovid MEDLINE(R) In-Process & Other Non-Indexed Citations and Ovid MEDLINE(R) < 1950 to present >

Searched via Ovid interface: 20 October 2009

1. sepsis/ (35,031)

2. sepsis.tw. (48,315)

3. shock, septic/ (15,639)

4. septic shock.tw. (10,675)

5. septicaemia.tw. (4752)

6. septicemia.tw. (9308)

7. 1 or 2 or 3 or 4 or 5 or 6 (91,334)

8. quality adjusted life year/ (4132)

9. quality adjusted life.tw. (3524)

10. (qaly$or qald$or qale$or qtime$).tw. (2923)

11. disability adjusted life.tw. (694)

12. daly$.tw. (734)

13. (sf6 or sf 6 or short form 6 or shortform 6 or sf six or sfsix or shortform six or short form six).tw. (1020)

14. (sf12 or sf 12 or short form 12 or shortform 12 or sf twelve or sftwelve or shortform twelve or short form twelve).tw. (1451)

15. (sf16 or sf 16 or short form 16 or shortform 16 or sf sixteen or sfsixteen or shortform sixteen or short form sixteen).tw. (20)

16. (sf20 or sf 20 or short form 20 or shortform 20 or sf twenty or sftwenty or shortform twenty or short form twenty).tw. (300)

17. (euroqol or euro qol or eq5d or eq 5d).tw. (1875)

18. (hql or hqol or h qol or hrqol or hr qol).tw. (4200)

19. (hye or hyes).tw. (50)

20. health$year$equivalent$.tw. (37)

21. health utilit$.tw. (660)

22. (hui or hui1 or hui2 or hui3).tw. (610)

23. disutili$.tw. (126)

24. rosser.tw. (63)

25. quality of wellbeing.tw. (3)

26. quality of well being.tw. (248)

27. qwb.tw. (137)

28. willingness to pay.tw. (1286)

29. standard gamble$.tw. (564)

30. time trade off.tw. (509)

31. time tradeoff.tw. (178)

32. tto.tw. (384)

33. or/8-32 (17,318)

34. 7 and 33 (52)

35. limit 34 to yr=“2004 -Current” (32)

36. limit 35 to english language (32)

37. exp animals/not humans/ (3,478,640)

38. 36 not 37 (32)

EMBASE

Database: EMBASE < 1980 to 2009 week 42 >

Searched via Ovid interface: 20 October 2009

1. sepsis/ (43,984)

2. sepsis.tw. (40,106)

3. septic shock/ (14,047)

4. septic shock.tw. (9145)

5. septicemia/ (9193)

6. septicaemia.tw. (3523)

7. septicemia.tw. (6151)

8. 1 or 2 or 3 or 4 or 5 or 6 or 7 (81,698)

9. quality adjusted life year/ (4481)

10. quality adjusted life.tw. (3029)

11. (qaly$or qald$or qale$or qtime$).tw. (2491)

12. disability adjusted life.tw. (503)

13. daly$.tw. (529)

14. (sf6 or sf 6 or short form 6 or shortform 6 or sf six or sfsix or shortform six or short form six).tw. (880)

15. (sf12 or sf 12 or short form 12 or shortform 12 or sf twelve or sftwelve or shortform twelve or short form twelve).tw. (1141)

16. (sf16 or sf 16 or short form 16 or shortform 16 or sf sixteen or sfsixteen or shortform sixteen or short form sixteen).tw. (14)

17. (sf20 or sf 20 or short form 20 or shortform 20 or sf twenty or sftwenty or shortform twenty or short form twenty).tw. (198)

18. (euroqol or euro qol or eq5d or eq 5d).tw. (1594)

19. (hql or hqol or h qol or hrqol or hr qol).tw. (3519)

20. (hye or hyes).tw. (32)

21. health$year$equivalent$.tw. (28)

22. health utilit$.tw. (566)

23. (hui or hui1 or hui2 or hui3).tw. (451)

24. disutili$.tw. (101)

25. rosser.tw. (55)

26. quality of wellbeing.tw. (5)

27. quality of well being.tw. (206)

28. qwb.tw. (119)

29. willingness to pay.tw. (1108)

30. standard gamble$.tw. (478)

31. time trade off.tw. (452)

32. time tradeoff.tw. (155)

33. tto.tw. (353)

34. or/9-33 (15,078)

35. 8 and 34 (90)

36. limit 35 to yr=“2004 -Current” (73)

37. limit 36 to english language (72)

38. exp animals/not humans/ (14,494)

39. 37 not 38 (72)