Birmingham and Lambeth Liver Evaluation Testing Strategies (BALLETS): a prospective cohort study

Introduction

Liver disease represents a major source of morbidity and mortality in the UK. 1 Abnormal liver function tests (LFTs) have been shown to be predictive not only of liver disease mortality, but also of more general causes of mortality. 2 LFTs are a good example of inexpensive tests (modern auto-analysers process large batches of samples using inexpensive reagents) that are frequently ordered as a ‘test of exclusion’ in patients with non-specific symptoms, such as tiredness or upper abdominal discomfort. The tests are also non-specific in the sense that none of the four to eight analytes included in the LFT panel points directly to a specific diagnosis, and many are not even specific to the liver. A doctor may order a laboratory test because a patient has features of a particular disease; for example, the gradual onset of jaundice in a user of injectable substances points to hepatitis C. The prior risk of hepatitis in such a person would be high: many positives would be true-positives. In most cases, however, LFTs are ordered without such a traceable link between symptoms and a specific diagnosis, for example when patients have vague symptoms or as part of the monitoring of patients with chronic diseases. Such tests are often offered as a type of insurance policy, but the prior risk of disease is low and the predictive value of LFTs is, a priori, likely to be low also. LFTs are interpreted by reference to population norms, rather than explicit calculus of the relative benefits and harms of false-positive and false-negative diagnoses. Many patients have a positive test, but it is not clear what proportion of these are true-positives, especially when the test result is only mildly abnormal. Review of the literature (see Previous research) shows that there is little evidence from large cohorts of people with abnormal LFT results to guide clinical actions when LFT results are mildly abnormal. The issue of how, or even whether, to investigate abnormal LFTs under various scenarios is not settled.

It is clear that a very large number of tests are ordered and abnormal results are common. The laboratory at University Hospital Birmingham received 67,182 requests for LFTs in 2003, from 83 general practitioner (GP) practices representing 210 GPs. Of these, 9779 (15%) led to an abnormal result in the sense that at least of one of the analytes on the LFT panel exceeded the reference range. As LFTs are inexpensive and easy to organise as one of the standard ‘blood tests’ in the GP's repertoire, their use has become widespread without careful study of their meaning in a general practice setting. As the meaning of the various combinations of possible test results and clinical features is unclear, different practitioners respond in different ways to the same test profile – the eclectic nature of practitioners' responses to the same scenarios has been well documented. 3

Most abnormal LFT results are false-positives. Thus, large numbers of follow-on tests and much anxiety can ensue if a low threshold is used to define abnormality. On the other hand, there are arguments to adopt a low threshold for subsequent evaluation, as LFTs have the ability to detect diseases when they are most treatable, for example by reducing overload in patients with metal storage diseases or by administering antiviral agents in those with chronic viral hepatitis. Furthermore, theory-based interventions designed to modify behaviour that leads to liver damage, while clearly far from a panacea, nevertheless produces worthwhile benefits in that some people adopt healthy lifestyles when they perceive that their health is threatened and that engaging in the recommended behaviour will reduce this threat. 4–6

The incidence of many liver diseases is rising, for example with migration from places with high rates of chronic hepatotoxic viral infection, and as a result of alcohol and calorie excess. Comorbidity is becoming more common as alcohol misuse and calorie excess unmask other diseases of the liver, such as haemochromatosis.

Thus, three interacting factors create an urgent need to better understand the clinical epidemiology of abnormal LFTs:

1. frequent use of these tests

2. lack of clarity about the meaning of the results

3. increasing treatability and rising incidence of liver diseases.

A number of authors have produced diagnostic algorithms for the investigation of people with abnormal LFTs. 7–12 These provide sensible advice – for example stressing the importance of taking a careful family history or of responding to tests that suggest obstructive biliary disease – but they do not provide a clear probabilistic basis for their reasoning. In particular, there is no scientific rationale for the widespread advice to repeat an abnormal LFT before conducting further tests. Green and Flamm13 state in their 2002 review of 1400 papers: ‘Unfortunately … there are no long term prospective studies to define the natural history of liver disease in patients with abnormal liver chemistries tests.’ They call for a substantial prospective study of a well-documented population given a standardised diagnostic work-up in general practice and then followed up for a period of time. It was this gap in the literature that the Birmingham and Lambeth Liver Evaluation Testing Strategies (BALLETS) study was designed to rectify.

Previous research

There is considerable literature on the laboratory measurement of analytes. Dufour et al. 14,15 carried out a systematic review of this topic in 2000. This review contains much useful information on biological variability and how it is affected by sex, age, race, use of the oral contraceptive pill (and other medicines), pregnancy, exercise, delay in analysis and time of day. The study also reviews the patterns of abnormality of each analyte given different diseases. A further systematic review that distilled 14,000 references was commissioned by the American Gastroenterology Association Clinical Practice Committee in 2002. 13 Again, most of the references describe probabilities of test results given various diseases, rather than the probabilities of the various diseases given test results. For example, Bonacini 16 describes ‘test results in people with cirrhosis due to chronic hepatitis infection’. Only a small proportion of articles report likelihood of disease by test result. Studies in this category tend to be based on hospital patients with serious abnormalities, such as ‘notably raised aspartate aminotransferase’17 or ‘requiring liver biopsy’. 18–20 Angulo et al. 21 investigated a remarkable 733 patients with non-alcoholic fatty liver disease (NAFLD) confirmed by liver biopsy to determine which features were associated with more serious disease, while Ekstedt et al. 22 followed up 129 patients with biopsy-confirmed NAFLD for a mean of 13 years and showed that the subgroup with ‘steatohepatitis’ had an increased risk of both cardiovascular and liver-related death compared with a reference population.

We updated the above review (Table 1) and selected studies that started with the LFT result and then followed the cohort, so as to provide the type of probability needed for decision-making. MEDLINE was interrogated, with limits placed on the overall search with respect to ‘humans’ and ‘publishing date post 1980’. Owing to the variety of nomenclature regarding LFTs a variety of search strings were used for this category. Search strings relating to abnormal LFT results included ‘liver function test’, ‘transaminases’, ‘alanine aminotransferases’, ‘aspartate aminotransferases’, ‘alkaline phosphatase’ and ‘gamma-glutamyltransferases’. The search was focused by using the limits of blood, analysis and metabolism. Despite the limits, these search strings retrieved over 35,000 references. The term ‘hepatitis’ was considered too narrow when attempting to find studies that followed up patients for a variety of diseases, so the more general term of ‘liver diseases’ was included, with limits of diagnosis, enzymology, epidemiology, mortality and virology, which retrieved around 8500 references. When these two search strategies were combined, 1448 papers were returned, the abstracts of which were read.

Eight studies were found that matched our requirement of following up patients who had experienced an abnormal LFT result. Two additional articles were selected from the references of relevant studies. As a result, to the best of our knowledge, there are only 10 studies for which a cohort of asymptomatic patients with abnormal LFTs was followed up (Table 2). However, one article was written in Korean (only the abstract was translated) and was excluded from our analysis.

Two of the remaining nine English-language papers described record linkage studies. One such study was based on the Korean insurance database, which was linked with death certificates. 23 This study reported that increased alanine aminotransferase (ALT), even within the upper end of the normal range, was associated with eventual death from liver disease. A study carried out in Scotland linked general practice and hospital databases. 2,24 However, this was a retrospective study so a full liver screen was not conducted and follow-up was for a median of only 4 years, whereas many diseases, including chronic viral hepatitis, have much longer prodromal periods. 25

The other seven studies were prospective cohort studies, based on testing asymptomatic members of the general population. The famous Dionysos study,26 based on three analytes from the LFT panel, is included among these. In this study, an impressive 6917 citizens from two communities in northern Italy were screened. Although the authors tested for viral hepatitis all of those in whom the LFT result was abnormal (n = 1473), and among whom they found a prevalence rate of 2.4%, the main aim of their study was to determine the effect of alcohol and diet on LFTs. Testing for viral hepatitis was used as a method of excluding causes of liver damage other than their topic of interest, so in-depth analysis on how viral hepatitis affected the pattern of LFTs was not published. Another Italian study, by Pendino et al. ,27 screened 1645 inhabitants from a town in southern Italy, with both a LFT [ALT, aspartate aminotransferase (AST) and gamma-glutamyltransferase (GGT)] and viral screen. 27 The prevalence of viral hepatitis is much higher in this region because of a significant immigrant population, and the authors performed a more extensive analysis of the impact of viral hepatitis on LFTs. Of the 319 (19.4%) individuals in whom LFT results were abnormal, nearly 18% were infected with viral hepatitis. However, the LFT failed to detect 34 (37%) of the 92 cases of viral hepatitis present in the community. Perhaps the most comprehensive prospective analysis looking at the effect of viral hepatitis on individual analytes was carried out on a population of Japanese office workers. 28 The study used data from compulsory health checks, which included an ALT/AST/GGT panel along with certain additional tests, including a viral screen, which were added for study purposes. The authors found that ALT was the most sensitive of the three analytes used, detecting nearly half of cases of viral hepatitis, while being abnormal in 14% of the cohort (278 abnormal results in 1973 participants). The remaining four prospectively designed studies were carried out in general practice and were therefore closer in population terms to the BALLETS cohort. However, three of these are restricted to patients with persistently abnormal LFT results over a 6-month period,18,29,30 and one of these did not include a test for viral hepatitis. The final prospective study, by Whitehead et al.,17 was small and based on only one analyte.

After this review of the literature we concluded that no study has fully investigated a cohort of patients in primary care with an abnormal LFT result (from the full LFT panel) and no obvious or known liver disease. BALLETS is thus the first study to test the validity of the various strategies that a GP could use to make a diagnosis in patients with abnormal LFTs. The BALLETS study was based on performing a full LFT panel of investigations to identify diseases such as chronic viral hepatitis and primary biliary cirrhosis (PBC) that could otherwise be identified only by follow-up lasting many decades. The study was therefore designed to look into, and ‘concertina’, the future. Patients were also followed up over 2 years to detect systemic diseases attacking the liver (e.g. disseminated cancer), to follow the progress of people with excess alcohol consumption and/or ‘fatty liver’ on ultrasound and to ascertain the rates at which abnormal LFTs reverted to normal according to diagnostic category and type of analyte that was abnormal.

We also identified a relevant study by Kim et al. 31 This study prospectively followed a group of ‘healthy’ Korean factory workers, taking measurements of ALT, AST and GGT on at least two separate occasions. The full article was in Korean so we had access to the abstract only.

Purpose of integral pilot

Rather than follow convention and collect a full data set before setting out on the analysis it was decided to analyse data from the first practice to complete recruitment – the Hall Green Practice in Birmingham. This practice completed its recruitment at a point in time when recruitment in other practices was nascent or yet to begin. Analysis of the integral pilot was carried out as soon as the FU1 data became available, i.e. the integral pilot does not include the FU2 results.

The purposes of this pilot were threefold:

1. to ‘test the system’ by detecting incomplete data and exploring systematic failures so that remedial action could be taken where necessary

2. to compare patients entered in the study with those who might have been eligible but who were not entered in the study

3. to conduct a quality control study on the accuracy of ultrasound findings by reviewing a sample of images stored on tape.

Missing data

One hundred and sixty-one patients were entered in the study in the pilot practice. Two patients did not attend for the ultrasound examination and have been excluded from the pilot analysis. The following analyses all relate to the remaining 158 cases. Their age and sex distributions are shown in Table 8.

The index panel of LFT analytes was incomplete (i.e. not all of the eight results were available) for 26 out of the 158 patients and complete for 132 (84%) patients. The first follow-up panel of LFT analytes was not available in five cases and the panel was incomplete in 27 cases – thus complete data were available on the follow-up LFT panel for 126 out of the 158 (80%) patients. The full breakdown of the missing data is given in Table 9.

The missing data did not follow any anticipated pattern (see Table 9). One might have expected that if those cases for which all eight analytes required for study purposes had not been included then the five default analytes for this particular laboratory would have been measured. This would have resulted in bimodal distribution, with high peaks at eight and five analytes. On further enquiry, it transpired that the clerks who receive the request forms and enter the requests on computer do so with variable fidelity (for study patients as for routine patients). A programme of staff training was therefore put in place to try to reduce this problem. However, we were advised that with large numbers and high turnover of clerical staff in the laboratory, some remaining laboratory omissions were inevitable.

Comparison of patients who were and were not ‘recruited’

Some eligible patients declined to participate, but we became aware that many more were not invited to participate by their GPs. Furthermore, some GPs recruited many more patients than others. One possible explanation was a tendency to select patients with the more severely abnormal results for entry in the study. This tendency could have been motivated by a desire to obtain all of the ancillary tests inherent in entry in the BALLETS study while reducing the need for further attendances and testing among those at lower perceived risk. This could lead to bias if, even among cases with equal severity of abnormality, GPs were somehow identifying patients with the worst prognosis for inclusion in the study. This could result in exaggerated estimates of the risks associated with abnormal LFTs.

In order to shed light on this issue, we collected baseline data from all (195) eligible but non-entered patients for two calendar months – May and June 2006 – and compared them with 53 participating patients for those months. This epoch was selected on the grounds that it corresponded to the period of highest recruitment.

The 195 non-entered patients constituted two subgroups: 129 patients had simply not been invited by the GP, despite fulfilling all objective criteria of entry to the study, while the remaining 66 had declined to take part (Table 10a). These subgroups are broken down by age and sex in Table 10b. The mean age of the invited patients, 58.6 years, is somewhat higher than the mean age of not-invited patients, 54.1 years (p = 0.028, two-sided t-test). There was no significant age difference between ‘consenters’ and ‘refusers’ within the invited group (p = 0.766). Thus, the 53 patients in the study tended to be older than those outside it. To put this in perspective, 68% (40/59) of eligible 65- to 74-year-olds were invited to join compared with 31% (22/72) of eligible patients under 45 years.

By contrast, the sex distribution was stable across all subgroups.

Abnormalities in the index LFTs for these 195 patients are analysed in Tables 11 and 12. The proportion of patients with abnormal GGT was higher (p = 0.011) among those invited to join the study (73.7%) than among those not invited (58.1%). However, there is no evidence of preferential invitation associated with abnormality on any other analyte, nor with the presence of more than one abnormality in the index panel (see Table 11). However, there was an (unexplained) tendency for invited patients with abnormal globulin to decline to participate (p = 0.002). Otherwise we found no evidence of recruitment bias.

In the end, we were not able to exclude a degree of selection in patients entered. Some selection effects associated with age and GGT abnormality are suggested. Individual differences in how eligibility criteria are applied are inevitable in a large and busy practice and we cannot exclude a degree of bias owing to hidden confounders. If clinicians selected a group of patients with significantly higher prior risk, then it is possible that the study will somewhat overstate the association between mildly abnormal test results and the various disease end points. All we really know is that some apparently eligible patients were not invited by their GP to participate. This may be because of a purposive decision to exclude or because of some oversight. Such data are difficult to collect because doctors who are unwilling or too hard pressed to select patients for study entry are unlikely to go to the trouble of recording their reasons.

Quality control of ultrasound

In order to quality assure the liver imaging, the first (FU1) ultrasound images and paper reports of 50 randomly selected BALLETS patients were presented to the study radiologist, who was in complete agreement with the sonographer's findings in 34 out of the 50 cases. In the remaining 16 cases there was some ‘technical or relatively minor’ disagreement but ‘no serious clinical disagreement’ that might have altered clinical decision-making.

The process was repeated with 50 randomly selected FU2 ultrasound scans. The study radiologist agreed with the written report in 38 instances, and in the remaining 12 cases found that there was ‘technical or relatively minor’ disagreement, but, once again, no disagreement that would alter the clinical decision-making process.

Variables and data

Demographic and lifestyle information – including body mass index (BMI) and alcohol consumption in units per week – was coded using categorical variables. Six categories each were used for age and alcohol consumption and four for BMI. The details may be read from Table 14.

Concentrations of analytes in the LFT panels were recorded (see Table 4 for units) by the three individual laboratories. Laboratory-specific reference ranges, incorporating adjustments for age and sex (see Table 4), were used to categorise values as normal or abnormal. Thus, each LFT result was available in two forms: as a continuous variable (measured concentration) and as a dichotomous variable (normal/abnormal).

Liver fat on ultrasound was recorded on a four-point ordinal scale (normal, mild, moderate and severe). The condition ‘fatty liver on ultrasound’ was identified with the categories normal, mild and severe, and analysed as a binary variable.

Summaries of categorical variables (with percentages) are presented in tabular form. Summaries of analyte concentrations were expressed in terms of medians and quartiles.

Analysis of liver function test data

Diagnostic category and liver function test results

The relationship between diagnostic category and LFT results was investigated in two ways: (1) by adding diagnostic category to the explanatory models already derived and (2) by means of a multiple discriminant analysis. The discriminant analysis was carried out using a set patient-level variables and (logged) analyte concentrations that had been identified from a series of preliminary logistics regression analyses. The preliminary analyses involved separate stepwise logistic regressions designed to find significant predictors of individual disease categories. These predictors were then used in a multiple logistic discriminant (polytomous logistic regression) analysis between the separate diagnostic categories. The performance of the discriminant to distinguish between liver disease and a non-specific diagnosis was assessed using the area under a receiver operating characteristic (ROC) curve.

The stepwise element in the discriminant analysis described here was restricted to patients with a complete panel of LFTs at FU1. The diagnostic groups for serious liver disease were very small compared with the non-specific group, and further depleted in the complete case analysis. In order to make full use of the LFT data that were available from diseased patients, the analysis was repeated using a multiple imputation technique. Further details may be found in Chapter 4 (see Analysis of imputed data).

Fatty liver

Stepwise logistic regression was used to explore the relationship between fatty liver at FU1 and patient characteristics, including (logged) LFT results, BMI and alcohol consumption. In this analysis, linear and quadratic components for age and alcohol consumption were substituted for the categorical variable for ease of interpretation. These components relate to the ordered categories themselves rather than the raw data. Persistence of fatty liver from FU1 to FU2 was investigated by stepwise logistic regression, including fatty liver at FU1 as a predictor for fatty liver at FU2. The consequences for liver fat of a change in BMI within an individual subject were investigated by means of two ordinal regression analyses: (1) liver fat category at FU2 on liver fat category at FU1 with percentage change in BMI as a covariate and (2) numerical difference in liver fat category between FU1 and FU2 on percentage change in BMI. Change in alcohol units (on a square root scale) was incorporated as a covariate in these analyses.

Sample size

A main objective of the study was to investigate the connection between LFTs and serious liver disease. However, there were several diseases under consideration, and no single primary question on which to power the study [see Production of reference categories (categories of diagnostic groupings)]. In the original protocol, logistic regression methods were proposed to explicate the relationship between diagnostic group and analyte concentrations. Sample size calculations for such problems often focus on ‘events per variable’ rules, which, in this study, suggest that 5 to 10 positive diagnoses would be needed for each predictor variable. Here there are seven independent LFTs (given that total protein is the aggregate of two other analytes), suggesting that between 35 and 70 positive cases would be needed for an unadjusted analysis. The study was designed for 1500 patients, which gives a satisfactory number of events (60) assuming a prevalence of a positive diagnosis of 4%. In practice, 44 cases of serious liver disease were found. If serious liver disease is considered as a composite outcome, the events per variable approach suggests that a reliable analysis is possible, at least for the five non-protein analytes in the LFT panel together with a small number of covariates.

The ‘events per variable’ approach focuses purely on technical aspects of logistic regression estimates. In the original protocol, we also considered a novel alternative criterion based on the ability of a logistic discriminant to identify high-risk cases (i.e. patients with risk of disease higher than an acceptable threshold level). For this purpose a baseline level of acceptable risk was taken to be 2% of the average population prevalence. According to this approach, 1500 patients would be sufficient to estimate a logistic discriminant function with a 90% chance of flagging up any patient whose true risk was twice the acceptable baseline level. These calculations posited an average prevalence of 4% – close to that actually observed in the sample. In retrospect, it seems that the degree to which the risk is predictable from the LFT results was underestimated in the original calculations, suggesting that the true performance of the discriminant would exceed expectations. However, there does not seem to be any direct way to verify this, as the true risk profile remains an unknown function of LFT results. In any event, this approach is concerned only with case finding and attaches no penalty to false-positives.

The other principal statistical analysis is concerned with the incidence of fatty liver. For this the event rates are much higher than for serious liver disease, and the sample numbers required for a meaningful analysis are correspondingly less stringent.

Changes to protocol by section

• Section 2.5.6 The patient's perspective. Altered to describe the contents of psychology questionnaires.

• Section 3.2.1 Practices. Additional practices were recruited in order to improve and maintain recruitment to the study.

• Section 3.3.2 Formal enrolment in subsequent testing protocol: defining of the patient population and seeking consent. Changes to this section reflect alterations to the clinical process at each practice. The original clinical protocol was adapted to routine clinical practice.

• Section 3.3.4 Collection of patient information. Altered to indicate that psychology questionnaires would not be translated into languages other than English.

• Section 3.3.6 Long term follow-up. The 2-year follow-up phase in Birmingham was more complex than originally planned, including repeat USS, alcohol questionnaire and measurements. This section was altered to reflect changes to the process.

• Section 3.5.1 Broad aim. Altered to address the possibility of selection bias, which could occur when suitable patients declined to take part or when suitable patients were not selected by their GP to take part (see section 3.5).

• Section 3.7.1 Psychological pilot study. This pilot study was implemented to inform the development of psychological questionnaires for use in the main study. The study process was updated to indicate changes to measures and time points used for data collection.

• Section 5.1 Substudy: Cryogenic storage and later testing Approval was obtained to collect and store an anonymised blood sample from consenting Birmingham patients attending their 2-year (FU2) clinic appointment.

• Section 5.2 Substudy: A qualitative investigation into liver function test ordering behaviour of general practitioners involved in the BALLETS study. A substudy designed to examine the non-clinical motives behind a GP's decision to order an LFT.

• Section 5.3 Substudy: Follow-up of abnormal test results. In the course of the study some patients tested positive for some specific liver diseases, but many were not followed up according to the agreed algorithm for referral or further testing. Letters were prepared by the study hepatologist and chief investigator suggesting appropriate follow-up of individual study patients testing positive for particular diseases.

• Section 5.4 Substudy: Qualitative investigation exploring anecdotal and preliminary evidence that events associated with participation in the BALLETS study were motivational to patients with and without fatty liver. A random selection of study patients were interviewed, in response to anecdotal accounts from patients at 2-year follow-up clinics that they had implemented lifestyle changes following their first BALLETS clinic appointment.

Ethics committee approval for changes to the protocol

The main research ethics committee, St Thomas' Hospital Research Ethics Committee, gave favourable ethical opinion to the BALLETS study in April 2005.

During the recruitment and follow-up phases of the study, the St Thomas' Hospital Research Ethics Committee Modifications Subcommittee approved 10 substantial amendment applications for alterations to the study protocol and documentation. Detailed summaries of each amendment are provided in the appendices to the main report (Appendix 3, BALLETS study: summary of ethics and substantial amendment approval). All amendments were also approved by South Birmingham and Lambeth local research ethics/research and development committees.

Approval was sought for the recruitment of new study practices in Birmingham, and corresponding patient documentation, for two qualitative substudies, as described in Chapter 5 (see Psychology 1: effects of positive tests and Psychology 2: effects of results on behaviour), and for a more intensive 2-year follow-up phase for Birmingham patients, which included an additional USS and the cryogenic storage and later testing of cells and serum. In addition, approval was obtained for the study team to remind GPs by letter, of the need to follow up patients who tested positive for some specific liver diseases.

Diagnostic categories

Patients were categorised into three broad diagnostic groups, described more fully in Chapter 3 (Methods: main study). Categories 1 and 2 are the two broad ‘serious liver disease’ categories, which may be further subdivided into: category 1a (hepatitis B and C + other hepatocellular diseases). Category 1b (hepatic bile duct disease) and category 2 [other diseases (such as metastatic cancer)] affecting the liver. The remaining cases (category 3) are rather non-specific.

Ultrasound features

Sonography reports for the liver were obtained at FU1 for 1277 patients and at FU2 for 658 out of 1152 patients from Birmingham practices. Second ultrasound examinations were not performed in Lambeth (see Chapter 3, The 2-year follow-up visit). A four-point ordinal scale (normal, mild, moderate and severe) was used to describe liver fat on each occasion (see Chapter 3, Testing strategy for patients in the BALLETS study), with results summarised in Table 17.

The subsequent version of the table (Table 18) shows the persistence of the ultrasound diagnosis of fatty liver between the two epochs.

Liver function test panels

Summary of liver function test data

The blood samples were processed by three different laboratories (see Chapter 3, Selection of practices and patients), labelled 1–3. Each laboratory operates its own reference ranges for the detection of abnormality. Mostly the differences between laboratories were slight, but for one analyte (ALP) the reported results for laboratory 1 followed a markedly different distribution from the other two laboratories (see Appendix 2, Liver function test results by laboratory). Given the potential for differences in laboratory practice, the summary statistics in Table 24 have been computed separately for each laboratory (medians and quartiles).

TABLE 24a - Index LFTs

Q, quartile.

Analyte	Laboratory 1				Laboratory 2				Laboratory 3
	n	Q1	Median	Q3	n	Q1	Median	Q3	n	Q1	Median	Q3
ALT	898	22	32.5	50	79	31	48	65	137	22	35	55
AST	1049	23	28	38	79	29	40	53	30	27	33.5	44
Bilirubin	1049	6	9	13	79	6	8	11	137	9	12	20
ALP	1053	166	203	264	81	71	94	124	138	65	78	99
GGT	934	46	64	106	81	39	67	89	137	29	72	99
Albumin	1059	43	45	47	81	43	45	48	138	45	47	49
Globulin	863	27	29	32	71	28	31	34	43	26	29	32
Total protein	864	71	74	77	74	73	76	79	43	73	76	78

TABLE 24b - First follow-up LFTs (FU1)

Q, quartile.

Analyte	Laboratory 1				Laboratory 2				Laboratory 3
	n	Q1	Median	Q3	n	Q1	Median	Q3	n	Q1	Median	Q3
ALT	1021	22	30	46	84	29	45	60.5	129	20	31	44
AST	1018	22	27	36	82	28	35.5	47	112	25	32	39.5
Bilirubin	1018	6	9	13	84	6	8.5	11	131	7	10	16
ALP	1021	163	200	251	84	75	94.5	121	131	64	77	91
GGT	1027	38	59	99	88	34	58	92.5	128	27.5	52.5	92.5
Albumin	1039	44	46	48	84	43.5	46	48	131	45	47	49
Globulin	1015	27	30	33	84	28	31	33	115	26	29	31
Total protein	1027	73	76	79	88	73	77	80	120	73	76	78.5

TABLE 24c - Two-year follow-up LFTs (FU2)

Q, quartile.

Analyte	Laboratory 1				Laboratory 2				Laboratory 3
	n	Q1	Median	Q3	n	Q1	Median	Q3	n	Q1	Median	Q3
ALT	588	19	27	39	56	30.5	39	51	63	22	32	55
AST	640	21	26	32	28	24.5	34	39	56	26	34.5	52.5
Bilirubin	640	6	8	11	58	5	8	11	63	7	9	14
ALP	640	161	193	240	55	76	85	117	63	60	71	97
GGT	613	34	54	90	40	29	49.5	68.5	63	27	44	96
Albumin	659	44	46	48	56	44	45	47	63	46	47	49
Globulin	613	25	28	31	33	27	31	35	51	27	29	32
Total protein	623	71	74	77	33	73	77	81	51	74	76	78

The distribution of analyte concentrations is represented by the histograms in Figures 7 and 8. It is clear that the non-protein analytes exhibit substantial positive skewness. Accordingly, much of the analysis reported here deals with log-transformed LFT values, as discussed in Appendix 2 (see Liver function test results by laboratory). One advantage of this approach is that a multiplicative laboratory effect can be readily incorporated as an additive term in any linear model for LFTs. As necessary (e.g. in the stepwise analyses of Diagnostic value of liver function tests and Fatty liver on ultrasound, below) the log-transformed data have been explicitly standardised to zero mean and unit standard deviation (SD) within laboratories.

FIGURE 7.

Distribution of enzyme analytes from index LFTs (by laboratory).

FIGURE 8.

Distribution of protein analytes from index LFTs (by laboratory).

Analysis of unadjusted liver function test results

Patterns of abnormality and disease classes

The predictive value of index abnormality for individual analytes, and pairs of individual analytes, is investigated in Table 29. The table contains the percentage of patients in each liver disease category among those who have registered an abnormality in the analyte concerned, or on at least one member of a pair of analytes. It is interesting to compare these predictive values with prevalence of liver disease among the subsample of patients with a complete index panel of eight analytes. This is given in the first row of the table, and represents the PPV of the LFT panel.

TABLE 29 - Positive predictive values for specific disease categories (i.e percentage of patients with abnormalities who have disease) for individual analytes (and pairs of analytes) in the index panel

Analyte combination	Liver disease 1a	Liver disease 1b	Liver disease 2	Any liver disease: 1a + 1b + 2	No. abnormal
Complete panel	2.5	0.8	1.5	4.7	891
Single analytes
ALT	4.3	0.7	1.0	6.0	418
AST	5.6	1.2	0.8	7.7	248
Bilirubin	3.7	0.7	0.0	4.4	135
ALP	2.7	4.9	2.2	9.8	184
GGT	2.2	1.0	1.7	4.8	833
Albumin	3.6	0.0	0.0	3.6	28
Globulin	3.6	0.0	0.0	3.6	55
Total protein	4.3	0.0	0.0	4.3	94
Pairs of analytes
ALT or AST	4.6	1.0	0.8	6.3	505
ALT or bilirubin	4.2	0.8	0.8	5.7	527
ALT or ALP	4.1	1.8	1.2	7.1	564
ALT or GGT	2.7	0.8	1.5	5.0	963
ALT or albumin	4.3	0.7	0.9	5.9	442
ALT or globulin	4.3	0.6	0.9	5.8	464
ALT or total protein	3.9	0.6	0.8	5.3	486
AST or bilirubin	4.4	1.1	0.5	6.0	367
AST or ALP	4.1	2.3	1.3	7.6	394
AST or GGT	2.4	1.1	1.5	5.0	942
AST or albumin	5.5	1.1	0.7	7.3	274
AST or globulin	5.4	1.0	0.7	7.1	297
AST or total protein	4.6	0.9	0.6	6.1	328
Bilirubin or ALP	2.9	2.9	1.3	7.1	312
Bilirubin or GGT	2.3	1.0	1.5	4.7	930
Bilirubin or albumin	3.2	0.6	0.0	3.8	158
Bilirubin or globulin	3.2	0.5	0.0	3.7	187
Bilirubin or total protein	3.6	0.4	0.0	4.0	224
ALP or GGT	2.1	1.3	1.6	5.0	933
ALP or albumin	2.9	4.3	1.9	9.1	208
ALP or globulin	3.0	3.9	1.7	8.6	232
ALP or total protein	3.3	3.3	1.5	8.1	270
GGT or albumin	2.2	0.9	1.6	4.8	851
GGT or globulin	2.3	0.9	1.6	4.9	864
GGT or total protein	2.3	0.9	1.6	4.8	880
Albumin or globulin	2.5	0.0	0.0	2.5	79
Albumin or total protein	4.3	0.0	0.0	4.3	115
Globulin or total protein	4.4	0.0	0.0	4.4	113

One striking feature of the table is the poor predictive performance of GGT. It is outperformed by ALP in all disease categories and by ALT and AST for category 1a.

Positive predictive value may not be the most important criterion of diagnostic performance because a high value can be achieved despite missing a large fraction of cases present. However, it is the one measure that can be properly estimated from this study as our data relate to a comprehensive set of patients' abnormal LFTs.

True-positive rates can be estimated, but these can be applied only to a restricted population with an abnormal LFT panel. Of course, there may be cases of liver disease whose LFT panel happens to be normal. These estimates are shown in Table 30.

TABLE 30 - ‘True-positive rates’, i.e percentage of abnormal results within disease groupsa

Denominators here are the observed numbers in disease groups, but these are incomplete because the data relate only to abnormal LFT panels.

Analyte combination	Disease 1a (n = 32)	Disease 1b (n = 12)	Disease 2 (n = 15)	1a + 1b + 2 (n = 59)	No. of patients
Single analytes
ALT	66.7	37.5	28.6	51	1064
AST	48.3	25	14.3	34.5	1131
Bilirubin	16.1	8.3	0	10.5	1213
ALP	16.1	75	28.6	31.6	1220
GGT	64.3	100	100	80	1101
Albumin	3.1	0	0	1.7	1226
Globulin	8.7	0	0	4.4	951
Total protein	17.4	0	0	8.9	955
Pairs of analytes
ALT or AST	79.2	37.5	30.8	57.8	966
ALT or bilirubin	76.9	37.5	30.8	57.4	1054
ALT or ALP	76.9	75	46.2	68.1	1057
ALT or GGT	92.3	100	100	95.8	1051
ALT or albumin	70.4	37.5	28.6	53.1	1062
ALT or globulin	87	28.6	28.6	59.1	916
ALT or total protein	82.6	28.6	28.6	56.8	917
AST or bilirubin	51.7	33.3	14.3	38.2	1125
AST or ALP	55.2	75	35.7	54.5	1125
AST or GGT	76.9	100	100	87.2	1005
AST or albumin	51.7	25	14.3	36.4	1125
AST or globulin	54.5	12.5	15.4	34.9	927
AST or total protein	50	12.5	15.4	32.6	930
Bilirubin or ALP	29	75	28.6	38.6	1212
Bilirubin or GGT	70.4	100	100	83.3	1080
Bilirubin or albumin	16.1	8.3	0	10.5	1212
Bilirubin or globulin	18.2	12.5	0	11.6	939
Bilirubin or total protein	27.3	12.5	0	16.3	939

From Table 30 it appears that the combination of GGT and ALT gives the best overall disease coverage, identifying nearly 96% of all cases of serious liver disease.

The limits of abnormality used to define a positive diagnosis in Tables 29 and 30 are based on conventional reference ranges. These are determined so as to reflect the limits of “normal” physiology, but without regard to the impact of specific diseases. Because the BALLETS index panels include all patients with a positive LFT result (as conventionally determined), they can be used to investigate the effect of increasing the upper thresholds of abnormality on the numbers of positive diagnoses. If this approach is taken then the number of positives will inevitably fall, as some cases will be missed by the more stringent criterion. For an analyte which carries diagnostic information, it is to be expected that the ratio of true-positives to false-positives (the PPV) will increase as the threshold rises above the normal limit. This is investigated in Table 31 for thresholds set at twice the normal limit. A range of thresholds is considered in Figure 9: here the curves for ALT, AST and ALP lie consistently above the diagonal line, confirming that the ratio of true-positives to false-positives does in fact increase with increasing thresholds. Indeed the slopes of these curves rise sharply as they approach the origin, suggesting that a very high analyte concentration has very high predictive value for liver disease. For GGT the situation is less clear-cut. The ratio of true-positives to false-positives remains effectively constant as the threshold increases even to twice the conventional limit, rising only as it approaches a threefold increase. This observation may cast doubt on the value of GGT as a marker of liver disease in an unselected population, or at least suggest that its full diagnostic contribution is not captured by conventional reference ranges.

TABLE 31 - Effect on PPV of doubling the upper threshold of normalitya

Positives encompass any liver disease (1a + 1b + 2).

Analyte	Standard thresholds		Twice standard thresholds
	True-positives (n)	PPV (%)	True-positives (n)	PPV (%)
ALT	25	6.0	9	11.7
AST	19	7.7	4	14.3
Bilirubin	6	4.4	0	–
ALP	18	9.8	3	20.0
GGT	40	4.8	15	4.7

FIGURE 9.

The effect of increasing the threshold of abnormality for four analytes. Numbers of true-positives (i.e. patients in category 1 or category 2 with analyte concentration above the threshold) are plotted against numbers of false-positives, using thresholds set at fixed multiples of the current laboratory reference limit. Points are plotted at intervals of 0.1 up to twice the reference limit, and at 3, 4 and 5 times the limit. Points at 1, 1.5, 2 and 3 times the limit are labelled accordingly.

Diagnostic performance of alternative liver function test panels

As already noted (see Selection effects), the selected nature of the sample – confined, as it is, to subjects with at least one abnormal analyte – precludes direct estimation of absolute sensitivities and specificities of the LFT tests as markers of liver disease. However, the relative diagnostic performance of alternative LFT panels can be assessed by comparing the numbers of true-positives and false-positives in the index sample. These quantities are plotted in Figure 10 for each of the 255 (= 2⁸ – 1) possible LFT panels that can be constructed using eight analytes. In this analysis, ‘liver-related disease’ is defined broadly to include all group 1 and 2 diseases. The data are restricted to the 915 subjects with a complete set of index LFTs.

FIGURE 10.

Positive diagnoses from different LFT panels, split between the non-specific group (false-positives) and the pooled disease groups (true-positives). All 255 possible panels from the eight analytes are shown. Single analyte panels (open circles) and the complete panel of eight analytes (diamond) are identified. The frontier (solid circles joined by line segments) shows the best diagnostic performance that can be attained using the analytes ALP, ALP and GGT. Results from repeating a panel at follow-up if it is positive initially are indicated by the letter ‘R’ joined by an arrow to the initial panel.

One panel can be said to dominate another if it generates both more true-positives and fewer false-positives than its competitor. On this basis, the best candidates are those that are not dominated by any other panel. This set of candidates is well approximated by the panels on the frontier in Figure 10, which involve three analytes only: ALT, ALP and GGT. In theory, the overall ‘best’ panel would follow on specifying the trade-off between the value of a true-positive and the cost of a false-positive, and comparing this ratio with the slope of the line segments that go to make up the frontier in the figure. Using the three ‘frontier’ analytes only, the best panel would be determined as follows:

• Value–cost ratio < 10 It is best not to use routine LFTs for these patients.

• Value–cost ratio between 10 and 15 The best LFT panel is ALP alone.

• Value–cost ratio between 15 and 30 The best panel is ALP with ALT.

• Value–cost ratio > 30 The best panel is ALT with GGT.

The panels mentioned here delineate a plausible range of options, although, in practice, it may not be straightforward to specify an appropriate value–cost ratio.

The PPV of a panel depends on the ratio of true-positives to false-positives. For the 915 complete index panels that contribute to Figure 10 it ranges from 9.6% for ALP alone, through 7.0% (ALP and ALT) to 5.3% (ALT and GGT). The PPV of the eight-analyte panel is 4.8%.

The results above suggest that the functions of a routine LFT panel can be largely subsumed into just three analytes: ALT, ALP and GGT. Nevertheless, an LFT panel – however carefully chosen – will never function as a definitive diagnostic procedure for any particular pathology; symptomatic patients will continue to be monitored in a primary-care setting regardless of a ‘negative’ test result. In this context, a panel with high PPV is perhaps most useful to the clinician, particularly if a plausible biological interpretation of positive results can suggest a route towards a definitive diagnosis. For these reasons, we incline towards the recommendation that routine testing should be confined to ALP and ALT alone. Addition of GGT will certainly increase the number of positive results but will not necessarily help specify a future clinical pathway for the patient. As remarked above (see Figure 9) the PPV of an abnormal GGT result does not increase when a higher threshold of abnormality is used. This is not suggestive of a strong clinical effect.

Impact of patient characteristics on liver function test results

Disease categories

Over the course of the study, 59 patients were diagnosed with a serious liver-related condition. Of these, 32 had category 1a diseases (including 13 with hepatitis B or C), 12 had category 1b diseases, and 15 had category 2 diseases. The breakdown of these conditions is shown in Table 42. The remaining 1231 patients were classified as having no specific liver diagnosis. However, 53 of the non-specific group were not tested for at least one of hepatitis B and C – the most common of the serious diseases. These 53 patients were excluded from analyses that involve the diagnostic grouping, leaving 1178 patients in the non-specific group. A combination of non-specific and diagnostic categories yields 1237 patients.

Chronic viral hepatitis

Thirteen out of 1236 patients for whom results of both test were available (hepatitis B or hepatitis C test results were missing for 54 of 1290 patients; hepatitis B and hepatitis C results were missing in 38 cases) had chronic viral hepatitis: nine patients had hepatitis B and four had hepatitis C. The breakdown of the type of analyte that was abnormal in the index panel is shown in Table 43. In 10 out of these 13 cases, more than one analyte was abnormal. In eight cases, the ALT was abnormal, and there was no case in which AST was high but ALT was normal. In one case, only protein levels were abnormal and all the enzyme tests (ALT, AST, GGT and ALP) were normal. Fatty liver was present in six cases, a proportion that is similar to the overall rate of fatty liver in the complete data set (see Ultrasound features). One patient with hepatitis B has subsequently progressed to cirrhosis.

When ALT or AST levels were abnormal, the values tended to be more extreme in patients with viral hepatitis than in patients who did not have this disease (Table 44). The same trend was observed of all patients as a whole, not only those in whom the analyte was abnormal (data not shown). In nine participants either ALT or AST was abnormal. Country of origin was recorded in 1208 out of the 1236 patients in whom both viral hepatitis tests were undertaken: 107 were born in a medium- or high-risk country for hepatitis B or hepatitis C, according to World Health Organization (WHO) definitions, but 11 out of the 13 patients with chronic hepatitis originated from a medium- or high-risk country. None of the 13 patients admitted to use of intravenous drugs at any time.

Primary biliary cirrhosis and primary sclerosing cholangitis

Primary biliary cirrhosis was defined as a cholestatic blood picture with positive serology for anti-mitochondrial antibody (AMA). AMAs were positive in 13 BALLETS cases. Three were weakly positive, leaving 10 positive cases included as category 1b cases in the statistical analysis. In retrospect, one of those cases (case 7) may have been misclassified (Table 45). Nine patients had a diagnosis of PBC confirmed by liver specialist follow-up, with a strong predominance for female sex (8/9) and white race (9/9). The mean age was 69.1 years, with two-thirds being aged > 65 years. Other risk factors for liver disease, namely alcohol excess and obesity, were unremarkable in this PBC cohort with a mean BMI of 27.7 kg/m², of whom 100% consumed ≤ 6 units of alcohol per week on average. ALP was abnormal on index blood testing in all the identified female patients with PBC. The only exception was the male patient with a solitary GGT abnormality on index testing. The ALP remained abnormal on repeat blood testing at FU1 in seven individuals. The course of the disease is usually benign in patients detected by LFTs rather than features of cirrhosis. However, two cases in our sample had features of early cirrhosis on ultrasound.

In summary, if a GP identifies an incidental raised ALP (± GGT) in a white woman aged > 65 years, having excluded obesity and/or alcohol excess as causes of liver disease, it is likely to be cost-effective and clinically more intuitive to proceed straight to an AMA test rather than proceed to repeat tests or a full liver screen. An AMA test costs £8.01 (University Hospitals Birmingham NHS Foundation Trust, 2011) and the result should be available within 7 days, thus not delaying further clinical investigation if indicated. Two cases of PSC were detected in the study (Table 46).

Autoimmune hepatitis

Smooth muscle antibodies (SMAs) were positive in 47 cases (weakly positive in five of these). In two cases (Table 47) either the ALT or AST exceeded twice the ULN. These cases had not been followed up at the time of the report and the study hepatologist is reticent about making a firm diagnosis of this very rare disease in elderly patients. Moreover, such a diagnosis would use test results as both the topic of investigation and a diagnostic criterion thereby risking inclusion bias.

Haemochromatosis

Iron saturation exceeded 50% in 39 cases, and in eight of these it exceeded 80%. We obtained a haemochromatosis genotype for 27 cases with iron saturation > 50% during the 2-year follow-up.

These cases are summarised in Table 48, in which it can be seen that there are six cases of homozygous haemochromatosis (C282Y or H63D) and four compound heterozygote (C282Y + H63D) who may be classed as having haemochromatosis (category 1a disease). There were also two carrier heterozygotes (C282Y). Five of the six homozygous patients had iron saturations above 80%. In none of the compound heterozygote patients did iron saturation exceed this level. Three patients were deceased and one patient was no longer registered at this practice. Five patients did not attend follow-up clinic appointments. Three patients attended hospital liver clinic appointments but did not have HFE genotype results. Four out of the six homozygous cases had ferritin levels of > 1000 mg/dl and receive frequent venesection. Their families have been screened.

Wilson's disease

Four patients had abnormal caeruloplasmin levels (Table 49). Wilson's disease was excluded by 24-hour urine test in three patients. We reminded the GP of the remaining patient of the possible desirability of referral, but this does not seemed to have occurred.

Alpha-1 antitrypsin deficiency

Low A1AT levels were found in 47 patients. Thirty-seven have had phenotype testing and these were abnormal in three cases (Table 50). Cases 1 and 2 are under the care of a specialist. Case 3 has a lower risk phenotype (Pi MZ).

Alcoholic cirrhosis

There were five cases in which the hepatologist agreed (on the basis of the ultrasound picture and history) that the patient had alcoholic cirrhosis (with some overlap with haemochromatosis). A further case of hepatocellular cancer was detected in a patient who did not have hepatitis B, but biopsy confirmed diagnosis of non-alcoholic steatohepatitis (NASH).

Other diseases (category 2) involving the liver

Metastatic cancers in the liver, cancer of the pancreas/bile duct and hypothyroidism are the common diseases. We did not include incidental cancers (e.g. cancer of kidney) or gallstones confined to the gallbladder in this category.

Objective

In this section the possibility of using LFTs to predict diagnostic group is explored by means of a discriminant analysis incorporating appropriate adjustment for patient-level demographic and clinical variables. These comprise age, sex, BMI, ethnicity and country of birth. Alcoholic consumption is excluded; although it may be implicated in the onset of certain conditions, it is clearly a lifestyle variable and, moreover, is the one variable in the data set which has not been objectively determined. The aim here is to summarise clinical information in such a way as to inform diagnosis. The LFT panels used here were obtained at FU1.

The analysis concerns (and is restricted to) the patients diagnosed with category 1a and 1b liver disease, together with those in the non-specific diagnostic category. These groups comprise 1222 patients in all. For 14% of these patients some of the clinical and demographic data were missing. In order to make full use of the information available the complete case analysis was supplemented with an analysis using an imputation method to cope with missing observations.

Method (complete case analysis)

Results (complete case analysis)

Discrimination between diagnostic groups

From the results above, it appears that ALT, AST and ALP are the only LFT analytes that contribute independent discriminatory power to the problem of diagnosis. Accordingly, these were included in the discriminant analysis, alongside age, sex, BMI, ethnic group and country of birth. Four groups were retained in the analysis: non-specific; 1a, excluding hepatitis; hepatitis; and 1b. The potential of the method to distinguish between liver disease (categories 1a and 1b) and a non-specific diagnosis is summarised in Figure 11. This shows a ROC plot of the true-positive rate against the false-positive rate when the estimated probability of disease (i.e. probability of lying in either category 1a or category 1b) from a logistic discrimination analysis is used as a marker of disease.

FIGURE 11.

Error rates when the presence of liver disease (hepatitis, categories 1a, 1b) is assessed from a logistic discriminant analysis. The AUC = 0.839 [standard error (SE) 0.037].

According to this plot, when the threshold for a positive diagnosis is set so that the true-positive rate (sensitivity) is 90%, then more than half of the non-specific group will be misclassified as having liver disease (false-positive rate = 53%). This gives a pragmatic indication of discriminatory capability. In practice, the performance is likely to be worse than that depicted because the same data have been used both for estimation and assessment of the discriminant function.

Analysis of imputed data

Results (imputed data)

Similar stepwise procedures were followed as for the complete case analyses. The results are displayed in Tables 57–60. These may be compared with Tables 52–55. For disease category 1b there are eight complete cases, augmented to 12 in the imputation analysis. Nevertheless, the same variable (ALP) emerges as the only significant predictor in both sets of analyses. The situation in category 1a is less clear cut. Here there are 27 complete cases and 32 cases in the imputation analyses. ALT and AST emerge as the only LFTs that figure in any of the selected models, although never together in the same model. For the non-hepatitis category, category 1a (19 cases, 16 complete), BMI features in the imputation model, enhancing the discrimination achieved by ALT in the complete case analysis. In the imputation analysis, country of birth (in combination with ALT or AST) has superseded ethnicity and BMI (also with ALT or AST) in predicting hepatitis (13 cases, 11 complete).

TABLE 57 - Stepwise discrimination results for disease category 1a

Stepwise procedure	Variables retained	Pseudo R2	AUC
Backwards elimination	ALT, BMI	0.144	0.800
Forwards selection	AST, country of birth	0.135	0.755

TABLE 58 - Stepwise discrimination results for hepatitis B and C (amalgamated)

Stepwise procedure	Variables retained	Pseudo R2	AUC
Backwards elimination	ALT, country of birth	0.308	0.924
Forwards selection	AST, country of birth	0.302	0.894

TABLE 59 - Stepwise discrimination results for disease category 1b

Procedure	Variables retained	Pseudo R2	AUC
Backwards elimination	ALP	0.189	0.835
Forwards selection	ALP	0.189	0.835

TABLE 60 - Stepwise discrimination results for results for liver disease category 1a: excluding hepatitis

Procedure	Variables retained	Pseudo R2	AUC
Backwards elimination	ALT	0.075	0.758
Forwards selection	AST	0.096	0.756

The discriminant analysis described above for complete cases was repeated for the imputed samples, using ALT, AST and ALP alongside age, sex, BMI and country of birth. In contrast to the complete case analysis, ethnic group was omitted from further analysis as it does not feature in the imputed stepwise results. The implied diagnostic performance does not change substantially, although the trade-off between false-positive and true-positive rates (as measured by the AUC statistic) is marginally less favourable (Figure 12).

FIGURE 12.

Error rates when the presence of liver disease (categories 1a, 1b, hepatitis) is assessed from a logistic discriminant analysis using multiple imputation to handle incomplete cases. The AUC = 0.828. [Estimated AUC from combination of imputed samples using Rubin's method = 0.822, with standard error (SE) = 0.036.]

Discussion

The general conclusion appears to be that automatic diagnosis of serious conditions has little chance of success given the high false-positive rates. The diagnosis of viral hepatitis is the most promising possibility, with an estimated AUC of > 0.90. The power of the method relies heavily on the country of origin of the patient.

Among the LFT panel only three analytes make a significant contribution: ALP is the only useful predictor of disease category 1b; ALT and AST are both implicated in the diagnosis of category 1a diseases, though they appear to substitute for one another. It remains unclear which should be preferred if a choice has to be made.

Fatty liver in patients with abnormal liver function tests

The presence of an abnormal ALT or AST was associated with an increased likelihood of fatty liver, but the prevalence of fatty liver among patients with abnormal bilirubin or ALP was reduced. Results for GGT were of marginal significance and the protein analytes exhibited no clear effects (Table 61).

A detailed breakdown of the severity of fatty liver when different analytes (and pairs of analytes) are abnormal is given in Table 62. The PPVs refer to the performance of the LFTs in determining that the liver is in the mild, moderate or severe category on ultrasound. The aminotransferase enzymes (ALT and AST) have the highest PPVs.

The proportions of fatty livers according to BMI and alcohol consumption are shown in Figure 13 and Table 63. Among people with abnormal LFTs, the probability of fatty liver is over 64.6% if they are obese (BMI ≥ 30 kg/m²) drinkers, rising to 73.8% if the abnormality includes an abnormal ALT. However, there is a sizeable probability (31%) of fatty liver in participants who were neither (moderate to heavy) drinkers nor obese.

FIGURE 13.

Relationship between obesity, alcohol use and fatty liver. High alc, high alcohol consumption.

Patient characteristics associated with ultrasound diagnosis of fatty liver

The present study provides an opportunity for a detailed investigation of the impact of patient characteristics on fatty liver – especially the lifestyle factors of weight and alcohol consumption. LFT results may also be considered to see how well they discriminate between fatty and non-fatty livers without reference to ultrasound. The analysis was done using the FU1 results.

A logistic regression model was constructed for the presence of fatty liver on ultrasound using a backwards elimination approach. The explanatory variables considered were:

• sex

• age group (six categories)

• ethnic group (four categories)

• BMI (four categories)

• alcohol consumption (six categories).

These variables were entered into the model, together with all two-way interactions involving age or sex. Then all non-significant interaction terms (p > 0.05) were sequentially removed. At this stage, the intention had been to remove also non-significant main effects for variables not included in any surviving interaction (in practice, this eventuality did not arise). The method was repeated for an analysis of the ordinal category (‘severity’) for fatty liver using ordinal logistic regression.

The model was supplemented by LFTs (log transformed) obtained from a stepwise procedure involving both forwards and (potentially) backwards steps.

The backwards elimination and stepwise analyses were performed both for presence/absence of fatty liver (using ordinary logistic regression) and for severity (using ordinal logistic regression). The results, in terms of variables selected, were identical. Although several analytes are individually correlated with the presence of fatty liver, these correlations are subsumed into two analytes only – ALT and albumin. Once these two are entered into the model, no other analyte achieves predictive significance for the presence or severity of fatty liver.

Table 64 summarises the patient characteristic model from the logistic regression analysis. This model achieved a pseudo R² of 15.3% and an AUC measure of discrimination of 0.75. The ordinal regression results (not shown) are similar. It can be seen that BMI is by far the most important predictor here.

The model described above includes main effects for sex, age group (six categories), ethnic group (four categories), BMI (four categories) and alcohol consumption (six categories) together with the sex × alcohol interaction. The model may be simplified using linear and quadratic components for these terms as appropriate, leading to a more readily interpretable analysis.

Table 65 shows the effect on the deviance of sequentially replacing the categorical variables alcohol, age group and BMI by their linear and quadratic components. It appears from the p-values in the table that the fit of the model is not compromised by this process since the change in deviance is less than the change in degrees of freedom (df) in every case. Finally, it turns out that the interaction of sex with the quadratic component of alcohol (1 df) is not formally significant (p = 0.0697). It is convenient to omit this term from the final model, especially as its interpretation is not straightforward in any case.

The final (simplified) model (numbered ‘5’ in the table) predicts that the probability of having a fatty liver:

• has an inverted U-shaped relationship with age, reaching a maximum at around age 55 years

• increases with BMI

• increases with alcohol intake above 30 units per week, though with some variation between the sexes

• is less for patients of Asian origin (compared with white patients).

These features are shown graphically in Figures 14 and 15, which show the relationship between fatty liver, age and alcohol intake separately for males and females of normal BMI and BMI > 30 kg/m². It is apparent that fatty liver is more responsive to alcohol intake in females than in males. In males there is perhaps even a suggestion that alcohol may have a protective effect at low doses, a finding corroborated in the literature (see Discussion).

FIGURE 14.

Fatty liver, alcohol and age of patients with normal BMI (20–25 kg/m²).

FIGURE 15.

Fatty liver, alcohol and age of obese patients (BMI > 30).

Liver function tests were added to the model of Table 65 using a stepwise procedure. Only two LFTs (ALT and albumin) were retained and the results are summarised in Table 66. LFT is an important predictor here, second only to BMI. The pseudo R² = 22.1% and AUC = 0.802. These figures do not suggest that LFTs could furnish a reliable substitute for ultrasound for the determination of fatty liver.

Persistence of fatty liver from first to second follow-up

Thirteen patients were excluded from this analysis, as belonging to the serious disease categories (categories 1 and 2). This left 628 cases for analysis. LFTs, BMI and alcohol intake were all taken from the FU2 data. Logistic regression analyses were performed with FU2 fatty liver as the outcome variable. Fatty liver at FU1 was included as a predictor in the analyses.

Sparseness of data for some of the covariate combinations impeded the fitting of the above models to the second follow-up sonography data. (For example, there are no instances of fatty liver at FU2 among the eight subjects with BMI of < 20 kg/m².) To simplify the model fitting, age group (six categories) was replaced by linear and quadratic effects (2 df) and the category for BMI < 20 20 kg/m² was amalgamated with the base category (25 kg/m² ≤ BMI < 30 kg/m²). For ease of interpretation, units of alcohol were represented by the linear component of the six-level categorical variable.

The results (without LFTs) are summarised in Table 67 (pseudo R² = 29.8% and AUC = 0.845). As might be expected, the presence of fatty liver at FU1 is the most important predictor of fatty liver at FU2.

The results of adding ALT and albumin to the model are summarised in Table 68 (pseudo R² = 28.7% and AUC = 0.841). Fatty liver at FU1, BMI and ALT are the only variables that contribute independently to the chance of fatty liver at FU2. AST could stand in for ALT, but results in a marginally inferior fit. Albumin no longer features as a significant predictor either instead of ALT or in addition to it.

The effect of changes in body mass index and alcohol intake on fatty liver

It is clear from earlier sections that raised BMI is the most important risk factor for fatty liver. As BMI is partly determined by lifestyle and voluntary behaviour, it is of some interest to determine whether or not a change in BMI over the period of the study is associated with a concomitant change in fatty liver status for the individual patient. The analysis of this question is confined to the ‘non-specific’ diagnostic group. Table 69 suggests an association between even small reductions in BMI and improved liver fat.

The association was investigated by an ordinal logistic regression analysis, using a seven-level outcome variable, defined as the difference in the ordinal number of the liver fat category between FU1 and FU2, i.e. a measure of liver fat improvement ranging from −3 (representing a change from ‘normal’ to ‘severe’) to +3 (i.e. a change from ‘severe’ to ‘normal’). This outcome was regressed on percentage change in BMI, with a marginally significant result [p = 0.030, odds ratio (OR) = 0.76 (95% CI 0.60 to 0.97)] per 10 percentage points change in BMI.

Change in alcohol consumption (represented as a difference in units per week on the square root scale) was added to these models, though without any significant, or near significant, finding (Table 70).

These results are necessarily inconclusive, but furnish some confirmatory evidence for the benefits of weight loss on fatty liver.

It is of great interest to explore whether or not the finding of a fatty liver prompts weight loss. There was a non-significant change in BMI in the hypothesised direction. The main weight change in the fatty liver group is −0.4% and in the non-fatty liver group is 0.2% (p = 0.30).

Other ultrasound features

At FU1, the liver was abnormal in size in 58 cases (4.5%), and was large in all but two of these. Eight patients had a diagnosis of diffuse cirrhosis.

A focal lesion was found in 106 cases (8.2%), but in only 21 cases was the lesion suspicious (20) or obviously malignant (1). The gall bladder was identified in 1150 cases (90%) and gallstones were detected in 191 (17%) of those.

The extrahepatic bile ducts were dilated in 29 of the 1123 patients in whom they were seen (2.4%), and the mean ALP was higher in these cases (271 vs 203). The difference in ALP was even greater when 12 out of 1230 (0.98%) where the intrahepatic duct was dilated (364 vs 203).

Background

Chronic liver disease is often asymptomatic or associated with non-specific symptoms and its early diagnosis is usually through the use of blood-based LFTs, which are routinely requested in primary care. Although the result of an LFT might indicate serious liver pathology, an abnormal result is much more often a chance finding, with a predictive value of < 5%. In fact, as we pointed out earlier, the proportion of people who really benefit from an abnormal LFT result is much smaller than 5%. Most of the cases of haemochromatosis and PBC identified in BALLETS appeared to be progressing at a very slow rate and were likely to have lead times longer than patients' remaining lives. There must be great doubt about the benefits that would have arisen from identifying four cases of metastatic cancer. The 1% of patients with chronic viral hepatitis really did stand to benefit, but over 1300 positive test results is a considerable number when only 13 patients are to benefit. This ‘yield’ would be especially worrying if it was associated with anxiety sufficient to impair quality of life. On the other hand, long-term benefits might accrue if LFT results prompted people to adopt healthier lifestyles – a situation that could arise if LFT results were used to reinforce behaviour change advice in addition to their rather minimal diagnostic value. The psychological consequences of testing are therefore important. In this section we consider the effect of psychological testing on anxiety. We considered the effects of LFT results on behaviour in Chapter 4 and will do so again in this chapter (see Psychology 2: effects of results on behaviour).

A psychological evaluation was added to the main study to monitor any psychological harms created by reporting abnormal LFT test results to patients and informing them of their ultrasound results. Previous evaluations of the process of screening report negative effects on psychological outcomes including anxiety, depression and reduced quality of life in the short term, but little effect in the longer term. 43,44 Screening for potential liver disease, however, has not been investigated and this study therefore examined the psychological effects on patients.

Methods and rationale

Results

Not all patients were offered a psychological assessment questionnaire to complete, and some declined. Overall, 527 questionnaires were obtained following the index test (T₁). Two years later, T₂ questionnaires were returned by 596, of whom 243 had returned baseline questionnaires.

Table 71 shows the demographic and clinical characteristics of the 527 patients completing the T₁ questionnaire.

Baseline characteristics of patients completing T₁ (as shown in Table 72) were compared with those not completing the questionnaire: there were no significant differences. Similarly, there were no significant differences between this cohort and those responding to both T₁ and T₂ questionnaires except for age, with the latter being slightly older (mean age 59.5 vs 56.9 years; t-test, p < 0.01).

The impact over time of the report of an abnormal LFT was examined in those patients returning questionnaires at T₁ and T₂. Table 72 shows that both anxiety and worry declined significantly over the 2-year period; there was no change in self-reported health.

As the impact of the report of an abnormal LFT might have been amplified by the subsequent diagnosis of fatty liver following ultrasound, the change in emotional state was examined in those with and without a reported fatty liver. Figures 16 and 17 show similar declines in anxiety and worry over the 2 years, irrespective of the diagnosis of a fatty liver.

FIGURE 16.

Anxiety after the index test and 2 years later in subgroups with and without fatty liver.

FIGURE 17.

Worry after the index test and 2 years later in subgroups with and without fatty liver.

Discussion

The results of this study are in accord with previous research into the impact of screening, that screening might well raise initial anxieties and worries but these soon return to levels within the normal range. Previous research on the emotional impact of screening, however, has largely examined the impact of the screen on the bulk of patients, who are subsequently judged to be negative. In that context, initial anxiety might well be allayed by the reassurance of a negative test. In this study, however, the cohort recruited were those patients who had screened positive albeit with an indicator with predictive significance that was poorly calibrated but not likely to be high. As the GPs had to inform the patients of the results, it is likely that they were reassuring: LFT results were slightly raised, but this was probably of no serious clinical significance. In that sense, patients may have interpreted their abnormal blood test as within normal limits and therefore as a sort of negative.

However, fatty liver, which does suggest the early signs of liver disease, was diagnosed in one-third of patients. This diagnosis was reported to the patients after they had completed their baseline T₁ questionnaire. It might be expected that, if patients had been alarmed by a fatty liver diagnosis, this would have been apparent at 2 years. But again, whatever the initial concerns, these had clearly dissipated over time. Lastly, if sustained anxiety is a necessary ingredient of behaviour change, then this would suggest that the fatty liver diagnosis would not affect lifestyle to a material degree. If, on the other hand, anxiety is a necessary trigger for change which then becomes self-reinforcing then the initial anxiety may be sufficient. The initial levels of anxiety are slightly higher in those ‘with’ than in those ‘without’ a diagnosis of fatty liver.

In conclusion, this study confirms previous research showing that screening has no long-term emotional effects. Where it adds to these findings is that, when the screening result is ‘positive’ but surrounded by prognostic uncertainty, this too becomes normalised over time and has no long-term effect.

Background

The BALLETS study recruited 1300 patients from Birmingham and Lambeth practices. At the initial assessment 40% of study patients were found to have fatty liver on ultrasound. The recognised primary treatments for fatty liver are diet and regular exercise. 48

The literature on the subject suggests that the ‘working alliance’ between care provider and patient is important in adopting health behaviours. 49 This alliance is defined by the mutual agreement on goals and objectives and the extent of the emotional bond (liking or trust) between patient and provider. 50 In addition, care providers use a number of verbal compliance strategies in which the subtle use of language can help influence a patient's behaviour. 51 Recently, evidence emerged that moderate- and low-level lifestyle counselling interventions in patients with fatty livers are a practical and effective method of improving health. 52

At Birmingham follow-up clinics, where patients returned for a repeat USS, BMI and LFT, research nurses had positive anecdotal reports from patients regarding improved drinking, eating and exercise habits. Many reported having had an abnormal first ultrasound. The results were supported by preliminary analysis from the first 277 patients who were followed up at FU2. In the event, an association between a change in mass and a reversal of fatty liver to normal was confirmed in the final analysis (see Chapter 4, The effect of changes in body mass index and alcohol intake on fatty liver). We therefore sought an extension to conduct a qualitative study of BALLETS patients to better understand a possible modifying effect on behaviour of having an ultrasound showing fatty liver.

Methods and rationale

Based on the above evidence, we conducted a qualitative substudy to explore the patient's experience of participation in the BALLETS study with respect to the finding of fatty liver. We focused on the patient's perception of the results from the initial scan, how the results were imparted to the patient and whether the finding of a fatty liver led to making any lifestyle changes. Therefore, the main aim of this substudy was to understand the overall experience of taking part in the BALLETS study with special reference to the psychological impact of the finding of fatty liver on USS.

Forty patients who participated in the BALLETS study and attended for initial clinics (FU1) and follow-up clinics (FU2) were invited to be interviewed. These 40 patients were divided into four subgroups (Table 73) according to whether or not their BMI had reduced and whether or not their ultrasound showed fatty liver.

Patients were randomly selected across the four categories using BALLETS study participant ID numbers. Patients from all GP practices taking part in BALLETS were included. Because it was expected that not all patients would agree to take part in the substudy, 20 patients were randomly selected for each group. Patients were phoned by the substudy research associate, in list order. If patients could not be contacted or were unable to take part, the next patient on the list was invited, until the final sample of 40 was reached. Patients were sent information sheets after providing verbal consent. An appointment was made for the research associate to interview patients if they were in agreement.

All Birmingham, BALLETS study sonographers were invited to be interviewed to determine their opinions on the consultation process, the methods used to impart the results of the scan and possible implications.

Results

We interviewed 40 participants (see Methods and rationale) and five key themes emerged. These are described below and in Table 74. Five sonographers were also interviewed in order to gain a greater understanding of the consultation process.

Discussion

The growing commitment to patient involvement in research has been reflected by the expanding literature on the aims and core features of research from a patient's perceptive. There is, however, scant literature on the impact of research participation on patients, particularly regarding beneficial health effects resulting from behavioural and lifestyle changes.

Both compliance and adherence to lifestyle changes are influenced by a number of factors. Our initial hypothesis was that the results from the first ultrasound consultation acted as a powerful driver in motivating people towards improving behavioural and lifestyle factors, reflected in their change in BMI. Across the four groups, recall of participation in BALLETS was poor. Participants were uncertain when and how they received results, if at all. It was evident that in this context the impact of the consultation with the sonographer appeared to be minimal. Evidence elsewhere has indicated that even in the most serious of clinical cases patient recall of clinical information is poor:61 between 40% and 80% of medical information presented by health-care professionals is forgotten by patients. 62 A number of factors can contribute to this lack of recall, such as complicated medical terminology, educational status of the patient and the means by which the information is presented. 62 Existing literature indicates that participant recall of Central Office of Research Ethics Committees (COREC)-approved, informed consent information is poor, even among those with medical training. 63 The relevant details of BALLETS were contained within the information sheet, although the complicated constraints of a COREC-approved information form may have inhibited understanding and retention of this information and greater engagement with the study by BALLETS participants, including understanding the context of the consultation and implications of the results of the scan. 64

Individuals can be motivated to participate for several potentially interacting factors, including the likelihood of improved clinical care as a result of their involvement,65 and social influences such as a desire to please the practitioner. 66 As elsewhere, for the majority of those interviewed, participation was altruistically motivated. 67 The results were seen as of relevance to the study team and not to them as individuals.

Improving communication between patient and care provider, including adopting a less formal approach, can increase the likelihood of adherence to treatment and behavioural regimes. 68 The potentially more relaxed consultation between participant and sonographer exemplified a more informal discourse. Sonographers would impart information of low clinical impact during the scan, and as a result participants, even those in whom abnormalities were observed, reported that results were underplayed and as a consequence they felt no obligation to alter their behaviour.

Participants were aware of the requisites for a healthier lifestyle, some because of existing health conditions and others as a result of their participation in BALLETS. This capacity to obtain, process, and understand basic health information can then lead them to make appropriate health decisions and may account for the observed changes in liver status and BMI. The results within theme 5 indicate that there may be a relationship between being diagnosed with a fatty liver and an increase in awareness of healthy lifestyle factors. However, we did not find strong evidence that patients were powerfully motivated to change lifestyle by the finding of a fatty liver on ultrasound.

Background

The numbers of diagnostic tests used in public health systems are increasing in most countries69 (by 10% per annum in the UK over the last 3 years). 70 The proportion of tests originating from GPs is also increasing; requests from GPs accounted for 37.2% of biochemistry tests in 2002, compared with 41.7% in 2005. 70

Increases in the number of tests ordered could be because of a number of factors: an older population,71 increased range of tests available, increased expectations of patients and guidelines promoting multiple test use. 72 Increased testing inevitably produces more positive results, leading to knock-on investigations, adding further to the number of tests ordered. 71,73