An assessment of the cost effectiveness of magnetic resonance including diffusion-weighted imaging in patients with transient ischaemic attack and minor stroke. A systematic review, meta-analysis and economic evaluation

Inclusion/exclusion criteria

We modelled a typical population of patients presenting with suspected TIA/minor stroke including mimics based on existing data reflecting current demographic, pre-stroke medications, and stroke treatments using information from the literature and a large national stroke audit.

These data sources included:

• the North Edinburgh Stroke Study and Lothian Stroke Register (LSR, n = 2598, a hospital-based registry of all inpatients and outpatients with TIA and stroke presenting to the hospital between 1992 and 2000 with demographic information, imaging and laboratory investigations, past medical history, medications prior to and following the TIA/stroke, and quality-of-life data) and followed up to 4 years for recurrent stroke, cardiac events and death and used in two previous HTA reports5,26

• the Edinburgh Stroke Study (ESS, n = 1367, a hospital-based registry of all inpatients and outpatients with TIA and stroke presenting to the hospital between 2002 and 2005 with the same information except quality of life) and so far followed up to 4 years after the stroke111

• the Scottish Stroke Care Audit Study (SSCAS), a Scottish national routine collection of anonymised audit data from TIA and stroke services

• data from other studies of stroke and TIA with more detailed imaging, including a study comparing CT and MR scans in minor stroke between 1998 and 2001 (n = 230) conducted as part of a previous HTA-funded project31

• a study of patients with mild cortical and lacunar stroke between 2005 and 2007 (n = 250), all with MR DWI and T2* imaging114

• audit data of TIA and stroke mimics from stroke prevention services in Leicester (Eveson; obtained through NHS Improvement)

• data on quality of life at 6 months after stroke by functional outcome Oxford Handicap Score from the Clots in Legs Or sTockings after Stroke (CLOTS) trials. 115

Ethical arrangements

The use of existing anonymised data to populate the model and perform other analyses did not require additional ethics approvals. All data used in the proposed modelling were anonymised.

Proposed sample size

We modelled the effect of incremental change in proportions of TIA/minor stroke patients undergoing MRI for a typical UK population of 1000 patients. The existing data sets included several thousand patients with TIA, minor and major stroke, with initial clinical and imaging assessment, and data on functional status, stroke recurrence and death at 6 months, 1 year and later. We provide the results in the form of ‘events per a cohort of 1000 patients presenting to hospital with suspected TIA and minor stroke’.

Statistical analyses

Outcome measures

The primary outcome was the incremental cost-effectiveness of MR scanning compared with CT for the whole population. Secondary outcome measures included the number of strokes prevented per strategy, costs and QALYs. We tested the effect of substituting MR for CT in key subgroups by time to imaging, ABCD2 score, assuming that DWI-negative TIA patients did not have a TIA/minor stroke, and combining carotid and brain imaging in one examination. We tested the effect of changing the proportion of patients with TIA/minor stroke with a visible ischaemic lesion on ‘MR including DWI’ and CT brain scanning and the association with key clinical variables; whether adding information from brain scanning to clinical risk scoring improves prediction of future risk of stroke or death; the association between a positive or negative brain scan and risk factors for stroke, such as carotid stenosis; delays to diagnosis, delays to starting medical or surgical treatment if MR including DWI were to replace CT brain scanning; the number of disabling strokes prevented at 6 months, 1 and 5 years after TIA/minor stroke if ‘MR including DWI’ were to be used in most patients for a cohort of 1000 patients. We tested the effect of QALYs stratified by key patient groups. We will provide a range of options reflecting effect of using MR in different proportions of TIA/minor stroke patients to guide decision-making by health service purchasers.

A brief discussion of the specific aspects of each chapter is provided at the end of each chapter and a discussion of the whole project and main findings in Chapter 13.

Objectives

For the purpose of this HTA project we have conducted a systematic review considering all primary studies reporting the overall rate of stroke in TIA patients. Recent published systematic reviews on stroke risk after TIA (e.g. Wu and colleagues120 and Giles and Rothwell121) were used only as a source of relevant references. We followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines for systematic reviews of studies that evaluate health-care interventions to conduct this review. 123 We aimed to identify all published studies reporting the risk of recurrent stroke among patients with TIA and/or minor stroke irrespective of the clinical setting or type of study design.

Identification of studies

We searched indexed records that appeared in MEDLINE (Ovid) from January 1995 to November 2011. It is worth noting that for the purpose of this systematic review we combined the comprehensive literature searches we developed for Chapter 6 (DWI in patients with TIA and minor stroke) and for Chapter 4 (clinical prediction score and risk of stroke after TIA and minor stroke) of this assessment. The MEDLINE search strategy included both subject headings [medical subject headings (MeSH) terms] and text words for the target condition (e.g. stroke, TIA, minor stroke). We adapted the MEDLINE search to search EMBASE. In particular, we ‘translated’ the MEDLINE MeSH terms into the corresponding terms available in the Emtree vocabulary. We did not apply any language restrictions. The searches were initially run in November 2010 and updated in November 2011. Full details of the MEDLINE and the EMBASE search strategies are presented in Appendix 1. We imported all citations identified by the MEDLINE and EMBASE search strategies into the Reference Manager bibliographic database version 11 (ISI Research Software, San Francisco, CA, USA). We hand-searched all proceedings of the International Stroke Conference (2011) and the European Stroke Conference (2011, 2012). We also contacted experts in the field and perused the reference lists of all relevant articles to identify further published studies for possible inclusion in the review. Only full-text articles were deemed suitable for inclusion. 124

Inclusion/exclusion criteria

One review author (MB) examined the identified titles and abstracts and retrieved all potentially relevant citations in full. Full-text articles were assessed by two reviewer authors (MB, HM) and retained if they reported the number and proportion of patients with stroke recurrence at 7 days, 90 days, and/or > 90 days after index TIA. We analysed all studies regardless of whether they reported recurrent stroke at 7 days, 90 days or > 90 days. We excluded non-English articles for which a full-text translation was not available.

Quality assessment and data extraction

Two review authors (MB, HM) independently conducted data extraction and review the methodological quality of selected studies using the primary papers. Disagreements were resolved by discussion or referred to a third author (JMW) if necessary. We recorded data on study methods (e.g. setting, study design) characteristics of patients (first vs. recurrent TIA), and outcomes (stroke events at 7, 90 and > 90 days). We also collected information on the following methodological aspects, which we considered more likely to introduce potential biases: prospective compared with retrospective study design, data source for cohort identification (e.g. registries, databases, medical notes), patients’ selection criteria, definition of TIA and recurrent event, timing of clinical assessment, evaluating clinicians, and method of outcome ascertainment (active vs. passive).

Data synthesis

For each study we calculated the total number of patients with stroke recurrence after index event (first or recurrent TIA). The pooled proportion of TIA patients with stroke recurrence was calculated by means of a univariate random-effects meta-analysis with within-study variance modelled as binomial. Heterogeneity between studies was assessed visually by inspection of the forest plots and by calculating I²-statistics. 109 Data were analysed in R version 2.14.2 (The R Foundation for Statistical Computing, Vienna, Austria) (cran.r-project.org).

Number of included/excluded studies

The literature searches identified 11,389 citations. After initial screening of titles and abstracts we selected 248 citations for full-text retrieval. Two additional studies published in 2005 were included after hand-searching of reference lists of relevant studies, yielding a total of 250 studies. After full-text examination, 197 studies were excluded, leaving 53 studies that fulfilled the inclusion criteria. The main reason for exclusion was that the number of recurrent events was not clearly reported (Figure 5). We translated one report published in Spanish; all other included reports were originally published in English. We observed, as have others,125 that some studies come from the same TIA cohorts (i.e. Oxford, north Dublin, California, Massachusetts, Lleida and Paris cohorts) and we experienced difficulty in avoiding double counting of the same patient data appearing in multiple publications. Many studies were published from the same research groups and the same cohorts of patients were described in multiple reports – with some reports assessing more than one patient cohort (i.e. Oxford, north Dublin, California, Massachusetts, Lleida and Paris cohorts). Even although we attempted to exclude multiple publications we found it challenging to identify which cohort was assessed in which study, as precise information were somehow lacking. When a single study included separate cohorts (e.g. community-based TIA patients and patients from specialist unit) for the purpose of the analyses we considered the separate cohorts as coming from separate studies.

FIGURE 5.

Identification and selection of studies.

Characteristics of included studies

The methodological characteristics of selected studies are shown in Table 1.

The 53 studies varied in their primary objectives. Eleven were population-based studies including cohorts from the UK (Oxford), Ireland (Dublin), USA (Texas and Minnesota), Canada (Alberta) and northern Portugal. 9,21,61,119,126,127,132,140,144,147,151 Seventeen were ED-based studies,23,69,87,129,131,134–136,143,149,153,156–159,161,163 three were focused on TIA clinics (one included OXVASC participants),24,130,145 10 were hospital-based studies (the authors of which did not describe characteristics of the hospital unit),65,81,120,133,142,148,152,155,160,162 10 recruited patients admitted to a specialist unit [including stroke unit, neurological ward or highly-specialised neurovascular (NV) clinics],122,137–139,141,146,150,154,164,165 one focused on patients attending ED and TIA clinics (including OXVASC participants)62 and one128 recruited patients as part of a hospital-based clinical trial. The 53 studies included a total of 30,558 participants. The risk of recurrent stroke at 7 days was reported in 28 studies (30 cohorts) of 12,332 participants, 34 studies (35 cohorts) of 19,769 participants reported the stroke risk at 90 days and nine studies of 8699 participants reported the stroke risk > 90 days. For studies reporting the stroke risk > 90 days, the length of follow-up ranged from 6 months to 14 years.

The study design was prospective in 33 studies9,21,24,61,65,69,81,119,122,126,128,129,132–134,136,138,139,141–144,148,150–154,156,160,161,164,165 and retrospective in 19;23,87,120,127,130,135,137,140,145–147,149,155,157–159,162,163 one study62 included two cohorts recruited prospectively and retrospectively. Participants were included consecutively in 45 studies. The diagnosis of TIA was made by a neurologist or stroke physician in 36 studies,9,21,24,61,65,69,81,87,119,122,126,128,130,132–134,137,139–142,144–148,150–152,154,156,157,160,161,164,165 by an emergency medicine physician in 10 studies,23,127,129,131,135,136,143,149,153,159 and by both in four;62,158,162,163 in three studies120,138,155 this information was not reported.

With respect to the timing of patient assessment after the index event, in 10 studies TIA patients were assessed within 24 hours of symptoms onset,62,69,87,134–136,138,143,155,161 in six studies within 48 hours,137,139,141,142,150,165 in one study140 within 72 hours and in one study160 within 7 days. Another four studies9,61,132,144 reported that ‘patients were assessed as soon as possible after the event’ but did not give a time. One study62 (including two cohorts) assessed patients within 24 hours and ‘as soon as possible after the event’, another study120 assessed patients ‘immediately’ after the event, and in three studies65,128,145 the timing of assessment was beyond 7 days. The remaining 27 studies did not clearly provide this information (see Table 1). Four studies included patients with first-ever TIA, 34 included either first-ever or recurrent TIA, and in 15 studies the authors did not report this information.

The majority of studies (n = 46) used a time-based definition of TIA (see Table 1), one study145 recruited patients retrospectively from clinical records using the International Classification of Diseases (ICD), Ninth Edition, definition of TIA, one study157 included just patients with transient symptoms and DWI visible lesion, in five studies the information was not reported. One study128 included a highly selected sample of patients with TIA and carotid stenosis, and three studies65,141,155 excluded patients with specific conditions [i.e. cardioembolic events or blood coagulation disorders, AF and posterior circulation TIAs, respectively].

The ascertainment of stroke events after TIA was carried out by face-to-face patient assessment or telephone interviews in 25 studies,9,21,24,61,69,81,119,122,128,132–134,137–139,144,149–152,155,156,161,164,165 by medical records in 15 studies,23,62,87,120,126,127,129,130,130,135,136,140,143,145,157,159 by mixed methods involving face-to-face assessment plus another method (i.e. telephone interviews, medical records, letters) in 11 studies,65,131,146–148,153,154,158,160,162,163 and was not reported in two studies. 141,142

Main findings

Table 2 shows the proportion of patients with stroke recurrence reported in all studies at different time points. The number of events was 974 out of 12,332 patients for studies reporting stroke risk at 7 days, 1567 out of 19,803 patients for risk at 90 days, and 927 out of 8699 patients for risk > 90 days. The risk of stroke reported across studies ranged from 0.0% to 22.4% at 7 days, 0.6% to 23.7% at 90 days, and 4.7% to 27% at > 90 days. There was evidence of significant heterogeneity between studies; the I²-statistics were 96.4%, 96.3%, and 97.8% for stroke risk at 7, 90 and > 90 days, respectively. Using random-effects meta-analysis, the pooled risk of stroke was 5.2% (95% CI 3.9% to 5.9%) at 7 days, and 6.7% (95% CI 5.2% to 8.7%) at 90 days. For stroke risk at > 90 days the pooled estimate was 11.3% (95% CI 7.5% to 16.6%). Figures 6–8 show the forest plots for studies reporting the proportion of patients with stroke recurrence at 7, 90 and > 90 days after index TIA.

FIGURE 6.

Forest plot for studies reporting stroke recurrence at 7 days after index TIA. The Johnston paper62 reports results for two cohorts: (a) the California clinic and (b) the Oxford clinic. The Rothwell paper61 also reports results for two cohorts: (a) the OXVASC cohort and (b) the hospital clinic.

FIGURE 7.

Forest plot for studies reporting stroke recurrence at 90 days after index TIA. The Johnston paper62 reports results for two cohorts: (a) the California clinic and (b) the Oxford clinic.

FIGURE 8.

Forest plot for studies reporting stroke recurrence beyond 90 days after index TIA.

Weaknesses of the review

Overall, our findings are limited by the great variation between included studies, especially in terms of study design, time of assessment from symptoms onset, type of TIA considered, and evaluating clinicians. As there is no a validated instrument for the quality assessment of observational studies on stroke risk, we pre-defined the methodological aspects that we considered more likely to represent potential sources of bias. The results of the quality assessment were used in only a descriptive way and not to inform further statistical analyses (e.g. meta-regression).

With regard to our literature searches, we focused mainly on two major electronic databases (MEDLINE and EMBASE). We did not search additional electronic databases, such as the Bioscience Information Service (BIOSIS), the Latin American and Caribbean Health Sciences Literature (LILACS) or the Science Citation Index (SCI), first because the number and relevance of indexed journals in these databases is limited compared with those indexed in MEDLINE and EMBASE, and, second, we are confident we have enhanced the overall sensitivity of our literature searches by hand-searching all recent conference proceedings of two major international stroke conferences, contacting experts in the field and perusing the reference lists of all relevant selected studies. Owing to time and resources constraints, we avoid searching grey literature sources, which are known to be time consuming and not particularly well developed.

Objectives

We followed the PRISMA guidelines for systematic reviews of observational studies to conduct this review. 175 We aimed to identify all published studies in which ABCD2 score was used to predict risk of stroke among patients with TIA and/or minor stroke irrespective of the clinical setting or type of study design and that reported on the actual rate of recurrent stroke.

Identification of studies

We searched indexed records which appeared in MEDLINE (Ovid) from January 2005 to November 2011. The choice of this time period reflected the introduction of the ABCD prediction score into clinical practice (even although the ABCD2 score was published for the first time in 2007, we adopted an over-inclusive approach and searched the literature from 2005 when the ABCD clinical prediction score was originally developed). The MEDLINE search strategy included both subject headings (MeSH terms) and text words for the target condition (e.g. stroke, TIA, minor stroke) and the prediction score under consideration (ABCD). We did not apply any language restrictions. We adapted the MEDLINE search to search EMBASE. In particular, we ‘translated’ the MEDLINE MeSH terms into the corresponding terms available in the Emtree vocabulary. The searches were initially run in November 2010, and updated in November 2011. Full details of both the MEDLINE and the EMBASE search strategies are presented in Appendix 1. We imported all citations identified by the MEDLINE and EMBASE search strategies into the Reference Manager bibliographic database. We hand-searched all proceedings of the International Stroke Conference (2011) and the European Stroke Conference (2011, 2012). We also contacted experts in the field and perused the reference lists of all relevant articles to identify further published studies for possible inclusion in the review. Only full-text articles were deemed suitable for inclusion.

Inclusion/exclusion criteria

One review author (MB) examined the identified titles and abstracts, and retrieved all potentially relevant citations in full. Full-text articles were retained if they studied the application of the ABCD2 score in cohorts of TIA and minor stroke patients and reported early risk of stroke assessed at 7, 90 and > 90 days. Studies that did not consider the incidence of further stroke in TIA/minor stroke patients; assess patients by means of the ABCD2 score; classify patients at high and low risk according to the cut-off score of 4; or report original data were excluded.

Quality assessment and data extraction

Two review authors (MB, HM) independently conducted data extraction and review the methodological quality of selected studies. Disagreements were resolved by discussion or referred to a third author (JMW) if necessary. We recorded data on study methods (e.g. setting, study design) characteristics of patients (first vs. recurrent TIA) and outcomes (stroke events at 7, 90 and > 90 days). We also collected information on the following methodological aspects, which we considered more likely to introduce potential biases: prospective compared with retrospective study design, data source for cohort identification (e.g. registries, databases, medical notes), patients’ selection criteria, spectrum of disease severity, definition of TIA, timing of clinical assessment, evaluating clinicians, and method of outcome ascertainment (active vs. passive).

Data synthesis

For each study we calculated the total number of patients with an ABCD2 score of > 4 and < 4, and the proportion of patients in each ABCD2 dichotomised category suffering a recurrent stroke at 7, 90 and > 90 days. The pooled risks of stroke for ABCD2 score of > 4 and < 4 at 7, 90 and > 90 days were calculated by means of univariate random-effects meta-analyses with within-study variance modelled as binomial. The discriminative ability of the ABCD2 cut-off score of 4 for stroke risk at 7, 90 and > 90 days was evaluated either by bivariate ROC curve random-effects meta-analyses or by random-effects univariate meta-analyses. We analysed all studies regardless of whether they reported recurrent stroke at only 7 or 90 days, and also restricted the analysis to just those studies that reported recurrent stroke at both 7 and 90 days, the latter to reduce the impact of some methodological differences between studies. We calculated estimates of sensitivity, specificity, positive and negative predictive values for a hypothetical cohort of 1000 TIA patients and expressed the proportion with recurrent stroke per 1000 patients assessed to improve the clinical relevance of the results. We used the statistical software R version 2.14.2 for our analyses.

Number of included/excluded studies

The electronic searches identified 3406 citations. Of these, we considered 59 reports to be potentially relevant and retrieved the full-text articles for detailed assessment. Hand-searching of reference lists and of recent issues of Stroke journal (not yet indexed in MEDLINE) resulted in a further two reports to be included. We subsequently excluded 33 articles. The most common reason for exclusion was that the study did not provide data in sufficient detail to allow calculation of the ABCD2 score for above and below 4. Twenty-six studies published in 28 reports fulfilled our inclusions criteria (Figure 9). Thirteen were observational prospective cohort studies and 13 were observational retrospective cohort studies. The included studies ranged in size from 69 to 1679 patients, and the total number of TIA/minor stroke patients assessed was 12,586.

FIGURE 9.

Identification and selection of studies.

Characteristics of included studies

The methodological characteristics of the individual studies are shown in Table 4.

All studies used a time-based definition of TIA as opposed to the recently new proposed tissue-based definition. 176

Four studies assessed population-based cohorts,132,140,144,151 three studies hospital-based cohorts,142,155,160 nine studies patients from EDs,87,134,135,143,149,153,158,162,163 and 10 studies patients from specialist stroke or neurology units. 62,137–139,145,146,150,152,164,177

In 15 studies the ABCD2 score was derived directly by patient assessment, whereas in the remaining 11 studies it was retrospectively calculated from medical notes.

Timing of patient assessment after symptom onset varied across studies. In seven studies ,TIA patients were assessed within 24 hours of symptoms onset, in four studies within 48 hours, in one study within 72 hours, and in two studies within 7 days. Another three studies reported that ‘patients were assessed as soon as possible after the event’ but did not give a time, and another study stated that the median time from symptoms onset was 15 days. Eight studies did not clearly provide this information (see Table 4).

Most of the studies included only patients with a definite or confirmed TIA by a neurologist/stroke physician and excluded TIA mimics. TIA diagnosis was made by a neurologist in 14 studies, by an emergency medicine physician in six studies, initially by an emergency medicine physician and subsequently confirmed by a neurologist in two studies and by a stroke physician in another two studies; The status of the diagnosing physician was not reported in the remaining two studies.

Little information was available on patients with carotid stenosis or use of endarterectomy. Similarly, information on how many TIA patients received secondary prevention drugs was rarely reported in the included studies. Only five studies stated that approximately 45–50% of TIA patients were administered antithrombotic therapy,143,152,153,158,162 and another study that 15% of patients received aspirin. 149

Ascertainment of stroke events after TIA was carried out by face-to-face patient assessment or telephone interviews in 17 studies, by medical records in five studies, by mixed methods (i.e. both telephone interviews and medical records) in three studies, and was not clearly reported in one study (see Table 4). Six cohorts reported only stroke events at 7 days,87,132,138,140,142,144 seven studies at 90 days,134,135,145,153,162,164,177 10 studies reported stroke events both at 7 and 90 days,62,137,139,143,149–152,158,163 and three studies reported stroke events at > 90 days. 146,155,160

Main findings

Considering all 16 studies that reported stroke events at 7 days, the total number of TIA patients with an ABCD2 score of ≥ 4 was 4590/6920 (66%) and the total number of recurrent strokes at 7 days was 488. In contrast, the total number of TIA patients with an ABCD2 score of < 4 was 2330/6920 (34%) and the total number of recurrent strokes at 7 days was 76.

In the 19 studies reporting stroke events at 90 days, the total number of TIA patients with an ABCD2 score of ≥ 4 was 6294/9849 (64%), and the total number of recurrent strokes at 90 days was 544 (Table 5). In contrast, the total number of TIA patients with an ABCD2 score of < 4 was 3555/9849 (36%) and the total number of recurrent strokes at 90 days was 80.

In the three studies146,155,160 reporting stroke events at > 90 days, 736/1073 (69%) patients with an ABCD2 score of ≥ 4 experienced a total of 152 recurrent strokes. In contrast, 337/1073 (31%) TIA patients with ABCD2 score of < 4 had a total of 44 recurrent strokes.

The corresponding pooled risks of stroke for patients with ABCD2 score of > 4 and < 4 at 7, 90 and > 90 days are shown in Table 6.

The pooled risks of stroke for patients with an ABCD2 score of > 4 and < 4 at 7, 90 and > 90 days for all included studies are shown in Figures 10–12.

FIGURE 10.

ABCD2 score of ≥ 4 and risk prediction at 7 days: all included studies.

FIGURE 11.

ABCD2 score of ≥ 4 and risk prediction at 90 days: all included studies.

FIGURE 12.

ABCD2 score of ≥ 4 and risk prediction at > 90 days: all included studies.

Many studies presented their results as accuracy of stroke risk prediction expressed as the area under receiver operating characteristic curve (AUROC) analyses. For completeness, we summarise the published area under the curve (AUC) values and 95% CI for stroke risk at 7 and 90 days in Table 7. We list these in order of publication year as it might be expected that patients would have been assessed more rapidly after TIA/minor stroke in the more recent studies (owing to greater awareness of the importance of seeking medical attention rapidly, better organised clinics, faster services, fewer recurrent strokes missed) and that this might have had some impact on the early stroke recurrence rate (higher rates in more recent studies) and in turn on the performance of risk prediction scores. However, there is no obvious trend in either the 7- or 90-day stroke risk or in the performance of the ABCD2 score in more recent years compared with earlier years, even allowing for the republication of some earlier data sets among later meta-analytic studies.

Figures 10–12 show considerable heterogeneity between studies and wide CIs, which may reflect variation in study methods or different populations of patients contributing to estimates of recurrence at different times, as outlined in the results above, rather than true differences in stroke rates at 7 and 90 days or in times to assessment. A previous meta-analysis found that the predictive value for 7-day stroke risk was higher in studies using in-person or clinical records to calculate the score than in studies with retrospective calculations of the ABCD2 score from hospital records. 178

To reduce the impact of any between-study heterogeneity, we considered separately the 10 studies that provided data on stroke risk at both 7 and 90 days. 62,137,139,143,149–152,158,163 The study by Asimos and colleagues was published in two reports. 136,143 Even although the same cohort of patients was studied in both publications, data on risk of stroke at 7 days were reported in the 2010 publication, whereas data on risk of stroke at 90 days were reported in the 2009 publication. The number of recurrent events at 90 days was strangely low compared with that at 7 days (see Table 5). Therefore, we excluded this study from further analyses to avoid introducing conflicting or mismatched estimates and examined the remaining nine studies that considered risk of stroke at both 7 and 90 days in the exact same cohorts of patients. These nine studies included a total of 4291 patients. Five of the nine studies identified patients retrospectively from review of medical records or equivalent. These five studies included 2019 patients, 47% of the total data from studies providing information at both 7 and 90 days. In these nine studies, the total number of TIA patients with an ABCD2 score of ≥ 4 was 2719/4291 (63%), and the total number of recurrent strokes at 7 days and 90 days was 176 and 264, respectively. In contrast, the total number of TIA patients with an ABCD2 score of < 4 was 1572/4291 (37%), and the total number of recurrent strokes at 7 and 90 days was 29 and 44, respectively. The pooled stroke risk at 7 and 90 days for these nine studies is shown in Table 8 (these studies did not provide data on stroke risk after 90 days).

The pooled sensitivity and specificity estimates for the ABCD2 score of ≥ 4 for predicting risk of stroke at 7 and 90 days are shown in Table 9.

Forest plots for the proportion of TIA/minor stroke patients with recurrent stroke at 7 days and ABCD2 score of > 4 and < 4 are shown in Figures 13 and 14 respectively, whereas forest plots for the proportion of patients with recurrent stroke at 90 days and ABCD2 score of > 4 and < 4 are shown in Figures 15 and 16, respectively. Results from the nine studies that provide stroke risk at both 7 and 90 days (see Figures 13 and 15: ABCD2 ≥ 4 and stroke recurrences), show reduced heterogeneity compared with those from all identified studies (see Figures 10 and 11). This suggests that as yet undefined methods differences between studies were the most likely sources of the considerable heterogeneity between studies rather than true differences in rates of recurrent stroke after TIA.

FIGURE 13.

Proportion of patients with an ABCD2 score of > 4 who have a stroke at 7 days from the subset of nine studies that provided data at both 7 days and 90 days.

FIGURE 14.

Proportion of patients with an ABCD2 score of < 4 who have a stroke at 7 days from the subset of nine studies that provided data at both 7 days and 90 days.

FIGURE 15.

Proportion of patients with an ABCD2 score of > 4 who have a stroke at 90 days from the subset of nine studies that provided data at both 7 days and 90 days.

FIGURE 16.

Proportion of patients with an ABCD2 score of < 4 who have a stroke at 90 days from the subset of nine studies that provided data at both 7 days and 90 days.

ABCD2 and proportions of patients with tight carotid stenosis

We examined all of the data on associations between the proportion of patients with high or low ABCD2 scores and an ipsilateral tight carotid stenosis, a key risk factor for recurrent stroke, to determine how many patients with tight carotid stenosis might have their investigations delayed and be at increased risk of stroke through having a low ABCD2 score. Four studies provided some data on carotid stenosis by ABCD2 score. 132,151,180,181 Ipsilateral carotid stenosis of > 50% was more frequent in patients with ABCD2 score of ≥ 4 but was not negligible in patients with ABCD2 score of < 4. In the SOS-TIA study (SOS Transient Ischaemic Attack), which enrolled a total of 1176 TIAs, among 679 of patients with ABCD2 scores of < 4, 9%, 6% and 8% of patients, respectively, had carotid stenosis of > 50%, AF or some other high-risk factor requiring immediate attention, compared with 14%, 11% and 10% of patients, respectively, among 497 with ABCD2 score of ≥ 4. 180 Basically, one-fifth of patients with an ABCD2 score of < 4 had a serious risk factor and one-third of patients with an ABCD2 score of ≥ 4. 180 In a further analysis of the SOS-TIA data, the 90-day stroke rate was 3.4% in patients with ABCD2 score of ≥ 4, 3.9% in patients with ABCD2 score of < 4 and carotid stenosis or other key risk factors, and 0.4% in patients with ABCD2 score of < 4 without key risk factors. 164 In another study assessing 220 TIAs, 11%, 22% and 22%, respectively, of 55 patients with ABCD2 scores of < 4 had carotid stenosis of > 50%, AF or both, compared with 14%, 18% and 18%, respectively, of 165 patients with an ABCD2 score of ≥ 4. 132 In a cohort of 443 specialist-confirmed TIAs, 23% of 169 patients with ABCD2 scores of < 4 had carotid stenosis of > 50%, compared with 25% of 274 patients with ABCD2 score of ≥ 4. 151 Another study that assessed TIA patients with an ABCD2 score of ≥ 5, showed that the recurrent stroke rate at 90 days increased with increasing degree of carotid stenosis. 182 Similarly, in a cohort of 40 patients with specialist-confirmed carotid territory TIA, 15% of 233 patients with ABCD2 scores of < 4 had carotid stenosis of > 50%, compared with 13% of 507 patients with an ABCD2 score of ≥ 4. 181 In short, the ABCD2 score does not predict carotid stenosis.

ABCD2 scores in transient ischaemic attack/minor stroke mimics

There was little information on ABCD and related scores in all patients presenting to stroke clinics with suspected TIA or minor stroke (i.e. including mimics), as most of the above studies specifically excluded non-TIA/minor stroke patients by specialist neurological examination and did not always differentiate TIA from minor stroke. In a retrospective cohort of 3646 patients presenting to an outpatient TIA service, of whom 1769 (48.5%) did not have TIA, a positive association was found between low ABCD2 scores and the diagnosis of non-CV events; after dichotomisation at 0–1 or 0–2 scores, the positive predictive values of ABCD2 score for a non-CV diagnosis were 0.81 and 0.74, respectively. 72 The analysis of the AUC for use of the score in the diagnosis of non-CV patients suggested a reasonable accuracy [0.74 (95% CI 0.73 to 0.76)]. 72 In a prospective cohort (north Dublin TIA study), the diagnostic usefulness of ABCD2 score to distinguish TIA or minor stroke from non-CV events was also assessed in 594 patients. 183 Mean ABCD2 score decreased across diagnostic groups [4.9, standard deviation (SD) 1.4 for minor ischaemic stroke; 3.9, SD 1.5 for TIA and 2.9, SD 1.5 for non-CV, p < 0.00001], as well the proportion of patients with ABCD2 scores of ≥ 4 and ≥ 6, respectively. An ABCD2 score of ≥ 4 increased the likelihood of a final diagnosis of confirmed TIA [odds ratio (OR) 2.8; 95% CI 2.0 to 3.9] and minor stroke (OR 8.4; 95% CI 3.8 to 18.9) compared with non-CV diagnosis. Scores of ≥ 4 had 60.3% sensitivity and 64.6% specificity for discriminating TIA from non-CV diagnosis and 82.2% sensitivity and 64.6% specificity for discriminating minor stroke from non-CV diagnosis. However, significant proportions of patients with a CV event (18% of minor stroke and 39% of TIA patients) had ABCD2 scores of < 4, and 35% of those with non-CV diagnoses had an intermediate or high ABCD2 score (≥ 4). 183 In a retrospective review (medical records) of 713 patients seen in 16 northern California EDs (from March 1997 to February 1998) who received an initial diagnosis of TIA, 79 patients (10%) had a final diagnosis of TIA mimic and 29 of them (41%) had an ABCD2 score of ≥ 4. 135 This indicates that, by placing over-reliance on the ABCD2 score, many true TIAs or minor strokes (with lower scores) would be missed while a substantial proportion of TIA mimics (35–41%) would be rapidly investigated.

ABCD2 scores and prediction of number per 1000 with recurrent stroke

Based on all available data for the ABCD2 score and predicted risk, we calculated the number of recurrent strokes by low or high ABCD2 scores in a hypothetical cohort of 1000 patients with probable or definite TIA attending a stroke prevention clinic. This requires that all patients initially referred to the TIA clinic as possible TIA/minor stroke have been filtered out by clinical assessment, which would have to be performed by a stroke physician or neurologist, as occurred in all published studies. We restricted the analysis to data from the nine studies that provided data on stroke at both 7 and 90 days. Among the 1000 selected probable or definite TIA/minor stroke patients there would be 634 with an ABCD2 score of ≥ 4 and 366 with ABCD2 score of < 4, of whom 30 and 6, respectively would have a recurrent stroke within 7 days, and 52 and 10, respectively, would have a recurrent stroke by 90 days. These numbers would shift substantially downwards if the starting population were the actual unfiltered population of patients that are referred to stroke prevention services with a suspected TIA/minor stroke, including the large proportion (45%) of patients with a stroke mimic as documented in the literature summarised in Chapter 7 and as found in the survey of UK stroke prevention services summarised in Chapter 8. Even a theoretical analysis of these data is substantially hampered by lack of information on the distribution of ABCD2 scores in all patients presenting to stroke prevention services. However, assuming that the neurologist was still filtering the patients into those with mimics and those with probable or definite TIA, then of the 1000 patients presenting to the TIA clinic, about 450, would have a mimic, and 550 would have TIA or minor stroke. Among patients with TIA/minor stroke, 349/550 would have ABCD2 score of ≥ 4 and 201/550 a core of < 4, of whom 16 and 5, respectively, would have a recurrent stroke by 7 days, and 27 and 5 by 90 days. Among the 450 with a mimic, 35–41%135,183 (say 38%) or 171 would have an ABCD2 score of ≥ 4 and 279 a score of < 4, making a total of 520/1000 (52%) clinic attendees with an ABCD2 score of ≥ 4, and 480/1000 48% with a score of < 4.

Identification of studies

Application of systematic review methodology to reviews of prognostic factors or prediction rules is not straightforward. Current methodology focuses on the retrieval and appraisal of particular kinds of study, particularly RCTs,191 yet most prediction rules are not developed in the context of RCTs. RCTs are carefully indexed as such in online medical databases. However, studies investigating the predictors of recurrent stroke are not restricted to any particular study design, especially not to RCTs, making it difficult to limit the number of hits found by a systematic literature search in medical databases, such as MEDLINE, by a study design filter. A simple, non-exhaustive, search strategy (see Appendix 1) in MEDLINE to find studies about recurrent stroke yielded 16,556 hits – an unworkably large number, given the resources available. Even for formal clinical prediction rules it is far more challenging to create an efficient yet exhaustive search strategy than for RCTs. 192

We therefore used a citation search method. Starting with a routinely collected reasonably comprehensive series of papers collected by one of the authors (JMW) over the last 20 years and several key papers,61,147,166,184 one reviewer (FMC) checked their references for relevant articles and also used Web of Knowledge ISI (http://wokinfo.com) to find articles that had cited the key papers. If a new study met our inclusion criteria, we also checked its references and found the studies that had cited it. This process was carried on iteratively until no new papers were found. This method has the advantages of transparency, repeatability and practicality. It can identify relevant literature without requiring the resources needed to administer thousands of abstracts and hundreds of full-text articles that would be identified by an exhaustive literature search but that would ultimately be irrelevant.

Inclusion/exclusion criteria

Each study had to contain data that could be used to ascertain recurrent stroke risk using either specified clinical prediction rules (ABCD, ABCD2, SP-II, the ESRS and the Californian rule; the CIP rule was not included as it is the ABCD rule plus DWI) or a described but un-named statistical model linking clinical predictors with recurrent stroke, with CT brain imaging or carotid imaging by any modality in adult patients who had had a recent TIA or minor stroke. Studies that did not use brain CT/carotid imaging, or used only imaging to predict stroke, were not included. Studies using a highly selected subset, for example young patients only, were noted but not included in the review. Studies were also excluded unless the predicted outcome was recurrent stroke. Studies published prior to 1990 were also excluded owing to developments in imaging technology since that time.

Quality assessment and data extraction

We derived a critical appraisal tool using a framework suggested for prognosis studies (see Appendix 2). 193 We particularly wished to examine patient recruitment into studies and patient selection for imaging as potential sources of bias. A brief inspection of the literature showed that patients may have had probable, possible or definite TIAs, imaging may have been reported for the entire or a subset of the sample, and recurrent stroke outcome events reported for 7, 30 or 90 days. Two other important issues were the definition of outcome and any stroke prevention treatment the patients may have received. Outcome was often simply defined as recurrent stroke without any indication of the severity. Also any treatment given will have an effect on stroke outcomes.

The data extracted from each article included the following:

• number of patients

• setting [e.g. hospital specialist clinic vs. community-based or accident and emergency (A+E) department]

• percentage male

• percentage TIA and minor stroke (possible, probable, definite)

• recruitment procedure (e.g. consecutive recruitment)

• age (mean/median age plus range)

• length of follow-up

• if patients at high risk of stroke were treated and the treatment given

• timing of baseline imaging from symptom onset or presentation

• clinical prediction rule or details of the statistical model

• type of additional imaging

• definition of recurrent event

• definition of positive imaging

• method of ascertaining effect of adding imaging to the clinical prediction rule/statistical model

• the added value of imaging to the clinical prediction rule/statistical model (this last may be expressed in a number of ways, most simply as the changes in the numbers of patients correctly predicted to have or not have a recurrent stroke).

One reviewer (FMC) performed the critical appraisal and extracted the data; ambiguous papers were discussed with a second reviewer (JMW) to minimise appraisal or data extraction errors.

Data synthesis

We planned to conduct two meta-analyses of the CT and carotid stenosis data. Studies were grouped according to their outcome measure [AUC, hazard ratio (HR), relative risk (RR)] for separate analysis as different outcome measures generally cannot be combined. We used forest plots and calculated I²-statistics109 to assess heterogeneity. All analyses were done in R using the generic inverse variance method in the meta-package.

Additional individual patient data from Lothian cohort

We also had access to data from a large prospective observational study – the LSR – conducted between October 1991 and 1999 in the stroke service at the Western General Hospital (WGH), Edinburgh, the methods of which have been published previously. 37 For the stroke prediction analysis, we used data on all patients referred with TIA, amaurosis fugax (AFx), retinal artery occlusion (RAO) or stroke, outpatients and inpatients between October 1994 and December 1999. Patients were excluded if they had a primary intracerebral haemorrhage (PICH) on CT/MRI scanning or were diagnosed with major stroke (abnormal verbal score on Glasgow Coma Scale, could not walk, could not lift both arms at examination). All patients were seen by a stroke physician mostly as outpatients in a rapid access stroke prevention clinic that operated twice per week in addition to patients being seen in A+E and on the wards. The details of the current and past history, and examination (with particular emphasis on vascular risk factors), blood tests, chest radiography, electrocardiography (ECG), CT brain and carotid artery imaging were entered into a database, of the LSR. The diagnosis of definite, probable or possible stroke/TIA/retinal ischaemia and not stroke/TIA/retinal ischaemia were determined by a consultant stroke physician in discussion with a panel of expert stroke physicians, prospectively, held weekly and blind to all follow-up events. The stroke physician also assigned an OCSP stroke subtype total anterior circulation cortical stroke (TACS), partial anterior circulation cortical stroke (PACS), lacunar stroke (LACS) or posterior circulation stroke (POCS). 194 The CT scans were coded by an expert neuroradiologist for the presence of any visible, recent and ‘likely to be relevant to the recent symptoms’ infarct for any old stroke lesions and non-stroke lesions using a structured validated scoring method. 195 The rater was not blind to the clinical presenting symptoms and signs so as to determine if any visible lesion were likely to be relevant or not, but was blind to any recurrent stroke or other outcomes and to the presence of other risk factors including carotid stenosis. Carotid stenosis was measured by ultrasound, the imaging being performed by one of two trained operators (a radiographer or neuroradiologist) who recorded velocities in the internal, external and common carotid and vertebral arteries [internal carotid artery (ICA), external carotid artery, common carotid artery (CCA), vertebral artery, respectively] on both sides of the neck. 196 In addition, any visible stenosis was noted, including its location, the degree of luminal narrowing by calliper measurement of the residual and original lumen at the point of maximum stenosis, the CCA lumen, and the ICA lumen distal to the stenosis by calliper measurements on the ultrasound console. 197 The degree of stenosis was assigned according to the ECST measurement method,82 using the velocity readings and calliper measurements, a method that has been validated against intra-arterial angiography (IAA). 197 The appearance of the stenosing plaque (smooth surface and of uniform echoity indicating ‘quiescent plaque’ or irregular surface and mixed echoity indicating ‘active plaque’) was also recorded. 198 The ECST velocity and calliper measurements were also converted to stenosis, expressed using the NASCET method for comparison with other studies. 199 Patients with > 50% NASCET (> 70% ECST) stenosis in the symptomatic ICA underwent confirmatory imaging with IAA prior to 1998 and MRA or repeat Doppler ultrasound (DUS) by a blinded operator thereafter.

Patients were followed up at 6 months, 1 year and 2 years by a trained researcher using a postal/telephone questionnaire, blinded to initial presenting symptoms and imaging features, using validated methods, to identify whether they had had a further stroke, myocardial infarction (MI) or died (and if so from what), and to determine their functional status. Therefore, the data were representative of practice in a rapid-access, same-day investigation NV clinic that accepted unscreened referrals from general practitioners (GPs) and other hospitals in the geographical area of north Edinburgh.

In all patients with mild stroke [defined as non-disabling and usually equivalent to a PACS, LACS, or POCS syndrome according to the OCSP classification,194 or TIA, or eye ischaemic symptoms (AFx or RAO), including those who would be candidates for endarterectomy] we performed multivariate survival analyses (Cox proportional hazards). The final predictive model included:

• residual neurological signs (yes/no)

• carotid stenosis (deciles, ECST criteria)

• age (years)

• previous stroke (yes/no)

• ocular not cerebral symptoms (yes/no)

• coronary illness (yes/no)

• visible infarct (yes/no).

Coronary illness was defined as the presence of one or more of the following: angina, previous MI, ECG showing acute or old MI. Predicted outcome was non-fatal disabling or fatal stroke. The model was rerun with and without carotid stenosis and CT visible infarct:

• Model A included all of the above except carotid and brain imaging (clinical only).

• Model B included all of the above except brain imaging (clinical plus carotid).

• Model C included all of the above except carotid imaging (clinical plus CT).

• Model D included all of the above (clinical plus carotid plus CT).

The models were used to generate an estimated risk of non-fatal disabling or fatal stroke for each patient, expressed as a probability from 0 to 1. A cut-off of 0.045 was applied to these risks to divide the patients into those predicted to have a stroke and those not predicted to have a stroke. The figure of 0.0045 was the cut-off chosen in the original LSR analyses for outcomes at 1 year.

The variables tested included age (years), residual neurological signs (yes/no), previous stroke, ocular as opposed to brain symptoms, coronary disease, the degree of ipsilateral carotid stenosis, atheroma plaque appearance or visible (likely to be relevant) infarct on CT. We tested the association between these variables and fatal or non-fatal disabling stroke during follow-up. Data were analysed with SAS version 9.1.3 (SAS Institute Inc., Cary, NC, USA; www.sas.com).

Number of included/excluded studies

The citation search method produced 1116 papers after removing duplicates and papers published prior to 1990; among this were 117 potentially relevant studies that were considered for inclusion. Ninety-six studies were rejected for the following reasons: 36 did not use brain CT or carotid imaging to predict recurrent stroke, 15 did not have recurrent stroke as an outcome, the relevant data were not reported in 13 studies, and a further 32 were rejected for a variety of reasons. This left 21 studies. The study designs ranged from population-based epidemiological studies to RCTs.

One of these 21 studies was an individual patient data (IPD) meta-analysis, which provided estimates of accuracy of prediction of the ABCD2 and brain CT individually for seven of the included studies. 178 We extracted CT results for each of the seven studies, but did not use the meta-analysis further, for example it was not included in the critical appraisal process. One of these seven studies also provided CT results in the primary publication. 69 A further 11 studies provided CT data, i.e. there were 18 studies providing data on CT brain imaging in total. There were 10 studies with carotid imaging data, and another 12 studies that used large artery atherosclerosis (LAA) where carotid stenosis was a component as a predictor. We recorded the results of the LAA studies for interest, but, owing to the heterogeneous definitions of LAA, did not formally include them in the review. Six studies had data on both CT and carotid imaging (Table 10).

We were careful to avoid duplicate publication of the same patient data, as some studies produced several publications. Only one reference per study was listed in the included studies, where that reference gave the required CT or carotid results. Secondary publications of included studies are given in Table 11, and were used for extra information in the critical appraisal and data extraction stages, with the exception of the studies used in the IPD meta-analysis,184 where the studies contributing to the IPD meta-analysis and which provided additional information on imaging findings, are listed as included studies, even if their own original publications did not provide CT brain or carotid imaging results.

Characteristics of included studies

The papers were broadly similar in their critical appraisal results (Figures 17 and 18). However, it was notable that the carotid stenosis papers mostly did define carotid stenosis adequately, whereas more than half of the CT papers did not give a description of a positive CT scan.

FIGURE 17.

Results of critical appraisal of studies contributing CT data. Each item on the checklist was coded ‘Yes’, ‘No’ or ‘Unclear’, with ‘Yes’ indicating better reporting or methodology (see Appendix 2). Note: item 5 is not included as it refers to carotid stenosis imaging.

FIGURE 18.

Results of critical appraisal of studies contributing carotid stenosis data. Each item on the checklist was coded ‘Yes’, ‘No’ or ‘Unclear’, with ‘Yes’ indicating better reporting or methodology (see Appendix 2). Note: item 4 is not included as it refers to CT imaging.

Many papers lacked details on exactly how the patients were recruited and, in particular, whether or not they were a consecutive sample, although the studies by Bonifati and colleagues,158 Cancelli and colleagues,200 and Sheehan and colleagues151 stressed the efforts made by the authors to identify all possible patients, and Sciolla and Melis69 and Purroy and colleagues150 explicitly stated that their sample was consecutively recruited. In addition, a few references had obtained data from RCTs (Table 12). 60,64,202,203

A more important point may be the various eligibility criteria used by the individual studies with regard to TIA status – some studies recruited patients if the emergency doctor suspected a TIA, whereas others recruited only patients with ischaemic stroke confirmed by a neurologist. The difference between these patient groups is likely to be a key factor in the rates of and prediction of recurrent stroke. However, the definitions of a TIA from study to study were generally very similar; a TIA was often defined as focal neurological symptoms lasting < 24 hours, thought to be of vascular origin (Table 13 and see Figures 17 and 18).

Most studies recruited patients from hospital settings, although three studies described themselves as population based. 61,151,208 The proportion of male patients ranged from 40.4% to 67.6%, but one study reported results for women only. 64 The reported average ages and age ranges were also broadly similar (Table 14).

The reporting of secondary prevention therapy offered to patients generally listed the medical therapies without information on the proportion of patients receiving them, or speed of administration. Some studies gave the numbers of patients undergoing carotid surgery, but none indicated how the removal of what is known to be a major source of stroke risk was accounted for in the prediction modelling (Table 15).

Many studies used the same definition of stroke, which was often the World Health Organization (WHO) definition of stroke or very similar (Table 16, and see Figures 17 and 18). Only one study specified that the recurrent stroke had to have a certain severity,143 two studies defined recurrent stroke as worsening or new symptoms201,202 and one study collected information on whether or not the recurrent stroke was disabling, but did not use this information in the analyses. 204

The definition of carotid stenosis varied from study to study. Three studies stated that the stenosis was dichotomised at 50%, but did not say whether this meant 50% by the NASCET or ECST criteria,200,201,208 and others also categorised their data, for example into 0–49%, 50–69% and 70–99% by the NASCET criteria. 151 Carotid stenosis was most often measured by ultrasound.

The definition of a positive CT scan also varied from study to study. Some specified that the infarct had to be new or consistent with the patient’s symptoms. 122,143,203 Two studies said that, owing to the difficulty of distinguishing between old and new lesions, the appearance of any lesion meant that the scan was deemed positive. 69,165 All of the other studies either did not report or inadequately reported their definition of a positive CT scan (see Table 16). The time of imaging from either symptom onset or presentation was recorded because the ability of CT to detect stroke lesions is heavily time dependent, particularly so in the first few hours after symptom onset. The timing varied from within 24 hours to within 1 month of symptom onset.

Main findings

Prediction of recurrent stroke

Seven of the CT studies’ results were provided by the IPD meta-analysis,184 expressed as the AUC of logistic regression models for the ABCD2 and ABCD2I clinical prediction scores. The total number of patients contributing CT data was 1368. The increase in the AUC for both 7- and 90-day recurrent stroke outcome was often modest, with a large overlap of CI between the AUCs of the two scores. The IPD study reported a summary 7-day AUC increase from 0.64 (95% CI 0.49 to 0.79) to 0.71 (95% CI 0.57 to 0.84), and a summary 90-day AUC increase from 0.69 (95% CI 0.65 to 0.74) to 0.75 (95% CI 0.70 to 0.80). 184

Unfortunately, only three studies,60,202,203 with a total of 4499 patients, provided data suitable for a meta-analysis of HRs. Each of the three studies used a statistical model rather than a simple score, with predictors not found in the other two studies. There was also evidence of significant heterogeneity demonstrated by both the forest plot (Figure 19) and a high I²-statistic of 78.5%. The summary HR was 1.74 (95% CI 0.82 to 3.70, p = 0.15); however, this is difficult to interpret beyond there being a positive though non-significant relationship between having a lesion found by CT and recurrent stroke. The difficulties derive from different patients groups (probable TIA202 vs. definite TIA)60,203 different definitions of lesion positive and outcome, and timing of imaging (Table 17) and different statistical models (Table 18).

FIGURE 19.

Forest plot of HRs of having a lesion found by brain CT in a multivariate model in the prediction of recurrent stroke. Summary HR is 1.74 (95% CI 0.82 to 3.70) from a random-effects meta-analysis of data from 4499 patients.

TABLE 17 - Primary study results for the addition of CT imaging to clinical variables for the prediction of recurrent stroke

DBP, diastolic blood pressure; SBP, systolic blood pressure.

Study	Model	Effect of addition of CT
Asimos and colleagues, 2010143	ABCD2 vs. ABCD2I (patients given an additional point for a positive CT scan)	AUC for logistic regression model increased from 0.68 (95% CI 0.54 to 0.83) to 0.86 (95% CI 0.76 to 0.95) for recurrent stroke at 7 days; results were identical for recurrent stroke at 90 days
Bonifati and colleagues, 2011158	ABCD2 with both CT and carotid stenosis added	Survival analysis (Cox proportional hazards) was used. CT (and carotid stenosis) remained significant (p = 0.02), whereas ABCD2 did not; only p-values reported
Coutts and colleagues, 2011201	Results for CT are available from a univariate analysis only	Survival analysis (Cox proportional hazards); HR was 0.8 (95% 0.3 to 2.1)
Cucchiara and colleagues, 2009139	ABCD2 vs. ABCD2I (patients given an additional point for a positive CT scan)	AUC for logistic regression model increased from 0.62 (95% CI 0.24 to 0.99) to 0.64 (0.23 to 1.00) for recurrent stroke at 7 days AUC for logistic regression model increased from 0.57 (95% 0.27 to 0.87) to 0.64 (0.32 to 0.95) for recurrent stroke at 90 days
Douglas and colleagues, 200335	ABCD2 vs. ABCD2I (patients given an additional point for a positive CT scan)	AUC for logistic regression model increased from 0.64 (95% CI 0.55 to 0.73) to 0.67 (95% CI 0.57 to 0.77) for recurrent stroke at 7 days AUC for logistic regression model increased from 0.70 (95% CI 0.62 to 0.78) to 0.72 (0.64 to 0.80) for recurrent stroke at 90 days
The Dutch TIA Trial Study Group, 1993202	Male sex, age > 65 years, dysarthria, multiple attacks, persisting > 6 weeks [minor strokes were also recruited], diabetes, intermittent claudication, haematocrit > 0.45 l/l, anteroseptal infarct, increased terminal P-wave	Survival analysis model (Cox proportional odds). Border zone infarct on CT, HR 3.5 (95% CI 2.0 to 6.0), any other infarct 1.6 (95% CI 1.2 to 2.0), white matter hypodensity 1.6 (95% CI 1.2 to 2.2)
Eliasziw and colleagues, 1995203	Age, TIA duration, ipsilateral stenosis, most recent event and history of TIA, hypertension, coronary artery disease, ipsilateral ulcerated plaque, occluded contralateral artery	Survival analysis (Cox proportional hazards); presence of infarct on CT was significant in univariate analysis (HR 2.31; 95% CI 1.04 to 5.16) but not in a multivariate model (HR 0.97; 95% CI 0.35 to 2.68)
Hankey and colleagues, 1992204	Sex, age, left ventricular hypertrophy, peripheral vascular disease, diabetes, carotid bruit, arcus senilis, cardiomegaly, ischaemic heart disease, AF, migraine, SBP, body mass index, number of cigarettes per day, alcohol units per day, carotid/vertebrobasilar presentation, AFx, residual neurological signs, low haematocrit, high haematocrit, infarct on CT on symptomatic side, infarct on CT for asymptomatic side, carotid stenosis on symptomatic side	Survival analysis (Cox proportional hazards). No numerical results reported. ‘. . . non-significant factors were . . . cranial CT evidence of infarction on the symptomatic and asymptomatic sides . . .’
Hier and colleagues, 1991205	Diabetes, diastolic hypertension (≥ 100 mmHg), prior stroke, infarct of unknown cause	Survival analysis (Kaplan–Meier); although an abnormal baseline CT scan was significant in univariate analysis (risk of stroke within 2 years, p = 0.03), it was not included in the multivariate model
Kernan and colleagues, 200064	Age, physical functioning, cognitive skills, > 80% carotid stenosis, prior TIAs or strokes, hypertension, coronary artery disease, myocardial dysfunction, AF, left ventricular hypotrophy, diabetes peripheral vascular disease	RR for stroke or death within 2 years of infarct on CT imaging adjusted for all other variables was 1.0. No CI provided
Lavallée and colleagues, 2007122	ABCD2 vs. ABCD2I (patients given an additional point for a positive CT scan)	AUC for logistic regression model increased from 0.77 (95% CI 0.58 to 0.95) to 0.85 (95% CI 0.72 to 0.98) for recurrent stroke at 7 days AUC for logistic regression model increased from 0.60 (95% CI 0.36 to 0.84) to 0.65 (95% CI 0.39 to 0.91) for recurrent stroke at 90 days
Purroy and colleagues, 2010150	ABCD, ABCD2, ABCDI, ABCD2I (patients given an additional point for a positive CT scan)	None of the models was found to be statistically significant predictors of ischaemic stroke at either 7 or 90 days. AUCs are 0.55 (95% 0.43 to 0.67) for ABCD, 0.57 (95% CI 0.45 to 0.69) for ABCD2, 0.53 (95% CI 0.41 to 0.65) for ABCDI, and 0.55 (95% CI 0.43 to 0.67) for ABCD2I. It is not clear from the text whether these figures refer to outcome at 7 or 90 days. Other analyses by the authors (not reported here) clearly demonstrate no statistically significant relationship between the models and recurrent stroke at 7 or 90 days
Purroy and colleagues, 2012165	ABCD, ABCD2, ABCDI, ABCD2I, ABCD3 (patients given an additional point for a positive CT scan, ABCD3 includes a point for prior TIA within 1 week of presenting event)	Method of analysis is described as ‘Cox regression analysis’, i.e. survival analysis; however, results are reported for specific time points, i.e. as if the method of analysis was logistic regression. Only ABCD3 was a significant predictor of recurrent stroke at either 7 or 90 days AUCs for stroke at 7 days: ABCD 0.57 (95% CI 0.46 to 0.68), ABCD2 0.56 (95% CI 0.45 to 0.66), ABCDI 0.56 (95% CI 0.44 to 0.67), ABCD2I 0.56 (95 % CI 0.45 to 0.67) and ABCD3 0.66 (95% CI 0.57 to 0.81) AUCs for stroke at 90 days: ABCD 0.55 (95% CI 0.46 to 0.64), ABCD2 0.55 (95% CI 0.46 to 0.64), ABCDI 0.56 (95% CI 0.45 to 0.67), ABCD2I 0.55 (95% CI 0.44 to 0.65) and ABCD3 0.61 (95% CI 0.52 to 0.70)
Rothwell and Warlow, 199960	Cerebral vs. ocular events, residual neurological signs after 7 days, diabetes, any ischaemic event in previous 2 months, number of events in previous 3 years, previous MI, degree of carotid stenosis, plaque surface irregularity, poststenotic collapse of the ICA, age, male sex, SBP, DBP, peripheral vascular disease, angina, left ventricular hypertrophy, occlusion of contralateral carotid artery	Survival analysis (Cox proportional hazards). Cerebral infarction on symptomatic side by CT was not found to be significant in a multivariate model, HR 1.32 (95% CI 0.88 to 1.96)
Rothwell and colleagues, 200561	ABCD2 vs. ABCD2I (patients given an additional point for a positive CT scan)	AUC for logistic regression model increased from 0.78 (95% CI 0.69 to 0.86) to 0.79 (95% CI 0.71 to 0.87) for recurrent stroke at 7 days AUC for logistic regression model increased from 0.67 (95% CI 0.58 to 0.76) to 0.71 (95% CI 0.63 to 0.8) for recurrent stroke at 90 days
Sciolla and Melis, 200869	ABCD2 vs. ABCD2I (patients given an additional point for a positive CT scan)	AUC for logistic regression model increased from 0.75 (95% CI 0.63 to 0.88) to 0.78 (95% CI 0.64 to 0.92) for recurrent stroke at 7 days AUC for logistic regression model changed from 0.76 (95% CI 0.66 to 0.86) to 0.76 (95% CI 0.65 to 0.88) for recurrent stroke at 90 days
Sheehan and colleagues, 2010151	ABCD2 vs. ABCD2I (patients given an additional point for a positive CT scan)	AUC for logistic regression model increased from 0.24 (95% CI 0.11 to 0.36) to 0.31 (95% CI 0.17 to 0.45) for recurrent stroke at 7 days AUC for logistic regression model increased from 0.61 (95% CI 0.38 to 0.85) to 0.70 (95% CI 0.50 to 0.90) for recurrent stroke at 90 days

TABLE 18 - Primary study results for the addition of carotid imaging to clinical variables for the prediction of recurrent stroke

DBP, diastolic blood pressure; NIHSS, National Institutes of Health Stroke Scale; SBP, systolic blood pressure.

Study	Model
Bonifati and colleagues, 2011158	ABCD2 with both CT and carotid stenosis added	Survival analysis (Cox proportional hazards) was used. Carotid stenosis (and CT) remained significant (p = 0.02), whereas ABCD2 did not; only p-values reported
Cancelli and colleagues, 2011200	It is not clear whether or not carotid stenosis was included in a multivariate model	Survival analysis (Kaplan–Meier) was used. Symptomatic internal carotid stenosis > 50% was not statistically significant, 0.136 vs. 0.103, p = 0.63
Coutts and colleagues, 2011201	Final model for stroke recurrence included blood glucose > 8 mmol/l and symptomatic internal carotid stenosis ≥ 50%	Survival analysis (Cox proportional hazards) in patients with an NIHSS score of ≤ 5, HR for carotid stenosis 5.6 (95% CI 2.0 to 15.6)
Eliasziw and colleagues, 1995203	Age, TIA duration, ipsilateral stenosis, most recent event and history of TIA, hypertension, coronary artery disease, ipsilateral ulcerated plaque, occluded contralateral artery	Survival analysis (Cox proportional hazards). Carotid stenosis was categorised to the nearest decile and was not found to be significant, HR 1.30 (95% CI 0.78 to 2.14), although this study excluded patients with < 70% stenosis
Hankey and colleagues, 1992204	Sex, age, left ventricular hypertrophy, peripheral vascular disease, diabetes, carotid bruit, arcus senilis, cardiomegaly, ischaemic heart disease, AF, migraine, SBP, body mass index, number of cigarettes per day, alcohol units per day, carotid/vertebrobasilar presentation, AFx, residual neurological signs, low haematocrit, high haematocrit, infarct on CT on symptomatic side, infarct on CT for asymptomatic side, carotid stenosis on symptomatic side	Survival analysis (Cox proportional hazards). No numerical results reported, although results were non-significant at the 5% level
Kernan and colleagues, 200064	Age, physical functioning, cognitive skills, > 80% carotid stenosis, prior TIAs or strokes, hypertension, coronary artery disease, myocardial dysfunction, AF, left ventricular hypotrophy, diabetes peripheral vascular disease	RR for stroke or death within 2 years for carotid stenosis > 80% in univariate analysis was 1.0; no CI or adjusted estimate provided
Marnane and colleagues, 2011208	Age, sex, hypertension, diabetes, smoking, AF, acute antiplatelet and statin treatment	Survival analysis (Cox proportional hazards). Carotid stenosis was the only independent predictor of stroke at 72 hours (HR 36.1; 95% CI 1.6 to 837.5); results also available for 7 days (HR 9.1; 95% CI 1.1 to 79.2) and 14 days (HR 4.6; 95% CI 0.9 to 22.8)
Purroy and colleagues, 2007182	Results available for a univariate analysis only	See Table 16 for definitions. Carotid stenosis absent, 5% risk of stroke, mild 25% risk of stroke, moderate, 11% risk of stroke, and severe 35% risk of stroke; p-value for trend < 0.0001
Rothwell and Warlow, 199960	Cerebral vs. ocular events, residual neurological signs after 7 days, diabetes, any ischaemic event in previous 2 months, number of events in previous 3 years, previous MI, degree of carotid stenosis, plaque surface irregularity, poststenotic collapse of the ICA, age, male sex, SBP, DBP, peripheral vascular disease, angina, left ventricular hypertrophy, occlusion of contralateral carotid artery	Survival analysis (Cox proportional hazards). Adjusted HR for carotid stenosis 1.25 (95% CI 1.18 to 1.40) in medically treated patients; carotid stenosis was no longer significant in the model for carotid endarterectomy patients
Sheehan and colleagues, 2010151	Age and sex	Survival analysis (Cox proportional hazards); results available for ≥ 50% stenosis (HR 2.65; 95% CI 1.28 to 5.20) and ≥ 70% (HR 3.23; 95% CI 1.45 to 7.21)

The possibility of bias must also be considered, given how small a subset of studies contributed to the meta-analysis.

Considering all CT studies, whether used in a meta-analysis or not, some found CT to be not significantly associated with recurrent stroke when adjusted for other predictors60,201,203–205 in survival analyses. Three studies also found CT to be non-significant with other analytic methods. 64,150,165 Only two studies found a significant relationship: one study that reported only p-values158 but another that gave fuller results202 (see Table 17).

The results for carotid stenosis provide stronger evidence of a relationship with recurrent stroke, with 6 out of 10 studies60,151,158,182,201,208 finding a significant relationship (see Table 18). One of the studies had not excluded patients with < 70% stenosis by the NASCET criteria,203 so could not provide an estimate of the risk of stroke between patients with lower degrees of stenosis compared with higher degrees.

Four of the studies provided data from 1944 patients suitable for a meta-analysis of HRs. 60,151,201,208 There was evidence of significant heterogeneity from both the forest plot (Figure 20) and the I²-statistic (81.9%), and the summary HR for carotid stenosis was 3.02 (95% CI 1.18 to 7.75, p = 0.02). Again, the interpretation of these summary statistics is challenging owing to different definitions of carotid stenosis and patient populations, and different statistical models used by the primary studies, but it does provide evidence of a link between carotid stenosis and recurrent stroke.

Perhaps the strongest evidence comes from an analysis of data from the ECST, in which carotid stenosis was a significant predictor of recurrent stroke in patients receiving medical treatment only, but not in those who underwent carotid endarterectomy. 60 When studies looked at the presence/absence of large artery atherosclerosis, which included carotid stenosis as a component, it was found to be a predictor of recurrent stroke at 90 days. 81,134,137,146,147,165,182,204–209

FIGURE 20.

Forest plot of HRs of carotid stenosis in a multivariate model in the prediction of recurrent stroke. Summary HR is 3.02 (95% CI 1.18 to 7.75) from a random-effects meta-analysis of data from 1944 patients.

The LSR registered 2940 patients with definite, probable or possible stroke, TIA, AFx or RAO, but 929 were excluded from the present analysis because of severe stroke or only possible cerebral or ocular ischaemic symptoms, leaving 2011 patients for inclusion in the analysis [1093 (54%) minor stroke, 828 (41%) TIA and the remainder had RAO]. Ninety-seven per cent of patients were followed up to 6 months, 90% to 12 months, and 62% to 24 months. The numbers suffering a stroke were 65, 85 and 82 by 6, 12 and 24 months, respectively. Of the recurrent strokes, 93% were ischaemic, and 48% were small cortical (PACS), 28% large cortical (TACS) and 9% were lacunar (LACS). Overall, 189/2011 (9%) patients had a tight (70–99%) symptomatic carotid stenosis and 86 (4%) patients had a carotid endarterectomy at a mean of 3 months after the presenting event (range 6 days to 1.5 years).

The analysis of the data from the LSR found both carotid stenosis and visible infarct on CT to be significantly associated with recurrent disabling or fatal stroke (Table 19). Unadjusted HRs for carotid stenosis by the NASCET criteria and visible infarct on CT were 1.006 (95% CI 1.003 to 1.008) and 2.45 (95% CI 1.73 to 3.46), respectively. The HR for carotid stenosis is small because it represents the change in risk for an increase of just one degree of stenosis. If carotid stenosis is dichotomised at 70% then the HR for patients above compared with below this cut-off is 2.42 (95% CI 1.63 to 3.61). When adjusted for residual neurological signs, age, previous stroke, brain compared with ocular symptoms, and coronary disease, the HRs for carotid stenosis and visible infarct on CT are 1.009 (95% CI 1.004 to 1.013) and 1.47 (95% CI 1.01 to 2.14), respectively. If carotid stenosis is dichotomised at the 70% cut-off, its adjusted HR is 2.15 (95% CI 1.42 to 3.23). The carotid stenosis HRs are also adjusted for CT and vice versa.

TABLE 19 - Results of survival analyses of the 2011 patients from the LSR cohort

Age, per year increase; carotid stenosis, per degree increase in stenosis; all other variables, yes/no.

Model	Predictor	HR	95% CI	p-value
Model A	Residual neurological signs	1.818	1.238 to 2.668	0.0023
	Age	1.031	1.014 to 1.048	0.0004
	Previous stroke	1.938	1.307 to 2.873	0.0010
	Ocular symptoms	0.449	0.195 to 1.036	0.0606
	Coronary disease	1.619	1.139 to 2.301	0.0072
Model B	Residual neurological signs	1.834	1.249 to 2.693	0.0020
	Carotid stenosis	1.009	1.005 to 1.013	< 0.0001
	Age	1.029	1.011 to 1.047	0.0013
	Previous stroke	1.824	1.228 to 2.711	0.0029
	Ocular symptoms	0.379	0.164 to 0.879	0.0238
	Coronary disease	1.524	1.072 to 2.168	0.0190
Model C	Residual neurological signs	1.577	1.050 to 2.368	0.0282
	Age	1.030	1.013 to 1.048	0.0006
	Previous stroke	1.999	1.348 to 2.967	0.0006
	Ocular symptoms	0.509	0.218 to 1.188	0.1184
	Coronary disease	1.648	1.159 to 2.343	0.0054
	Visible infarct	1.518	1.043 to 2.210	0.0293
Model D	Residual neurological signs	1.612	1.074 to 2.419	0.0211
	Carotid stenosis	1.009	1.004 to 1.013	< 0.0001
	Age	1.028	1.011 to 1.046	0.0016
	Previous stroke	1.898	1.276 to 2.823	0.0016
	Ocular symptoms	0.428	0.182 to 1.003	0.0509
	Coronary disease	1.552	1.091 to 2.209	0.0146
	Visible infarct	1.471	1.011 to 2.140	0.0437

None of the models A, B, C or D predicted that any patient would have a fatal or non-fatal but disabling stroke by 7 days. However, three patients did go on to have a non-fatal disabling stroke, another had a fatal stroke and there were also three other deaths within 7 days. All of these patients presented with minor stroke rather than TIA or RAO.

The results for 90 days and 2 years are given in Table 20. All of the models were better at correctly predicting a stroke outcome (fatal or non-fatal disabling stroke) at 2 years rather than 90 days. Conversely, all were better at predicting being alive and well at 90 days than 2 years.

TABLE 20 - Predicted and actual outcomes at 90 days and 2 years for the four predictive modelsa

Results for 7 days are not included, as no patients were predicted to have a stroke outcome at 7 days.

Outcome and time point	Model A	Model B	Model C	Model D
At 90 days
No. of correctly predicted fatal or disabling strokes	2	7	4	7
No. of stroke outcomes that were not predicted	26	21	24	21
No. alive and well, and predicted no recurrent strokes	1814	1798	1822	1800
At 2 years
No. of correctly predicted fatal or disabling strokes	84	80	82	81
No. of stroke outcomes that were not predicted	10	14	12	13
No. alive and well, and predicted no recurrent strokes	647	724	687	764

At 90 days there were 27 fatal and non-fatal disabling strokes, 20 patients with another non-stroke event (MI, vascular or non-vascular death) and 1963 patients with no further events. At 2 years there were 94 fatal and non-fatal disabling strokes, 103 patients with another non-stroke event and 1814 patients with no further events.

For all models, 91–92% of patients had a correctly predicted outcome at 90 days. This fell to 41–47% of patients when considering outcomes at 2 years. Model D (clinical variables, plus carotid and brain CT imaging) had the highest proportion of correctly predicted outcomes at 2 years (47%). It is obvious that the time after presenting event at which the outcome is predicted is crucial in the prediction of stroke.

Weaknesses of the review

Owing to the nature of the underlying studies, our methodology had to be adapted from the ‘gold standard’ methodology used for RCTs. In particular, the citation search method and the critical appraisal checklist are unvalidated, although both were extremely thorough and produced in as methodologically sound way as possible.

The results of the meta-analyses are based on a small subset of the available studies due to factors beyond our control. However, we include a larger sample than were included in recent meta-analyses on stroke risk prediction, for example, 1368 patients with CT184,210 compared with our 4499. The results are difficult to interpret given the differences between the primary studies, and there is also a possibility of bias, not least because the studies with non-significant results could not be included owing to a lack of data.

Weaknesses of the Lothian Stroke Register analysis

The data were collected in the 1990s and so represent older CT technologies than those that are in use today (DUS with colour has changed relatively little in practical terms). For diagnosis of carotid stenosis, the same velocity criteria and image measurement have been in use since the 1990s so the stenosis measures are relevant and valid. The CT technology has moved to thinner scan sections, which may increase detection of small lesions, but the contrast has not improved. The presence of a visible infarct on 1990s technology is as relevant as on scanners today, although lesions may have been missed that would have been visible on current technology. On the other hand, many of the literature reports concern patients recruited in a similar time period on similar technology therefore the LSR data are still highly relevant. There would have been more delay between symptom onset and patient assessment in the 1990s owing to greater awareness now of the need to seek medical attention quickly after TIA/minor stroke. Although the LSR took place in a rapid access clinic that ran several times weekly with same day CT and US and direct access from GPs, patients may not have contacted their GP quickly, thereby incurring delays prior to reaching medical attention. Similarly, this problem will have affected all of the studies that use data from the 1990s but should not make a material difference in terms of identifying the factors which most strongly predict recurrent vascular events.

Objectives

We performed this analysis according to the PRISMA guidelines for systematic reviews of observational studies. 175 Following an initial search for papers comparing MR with CT in the same TIA/minor stroke patients, we focused on identifying all published studies in which TIA and/or minor stroke patients were assessed by DWI irrespective of aims, type of study design, and clinical setting.

Identification of studies

We searched indexed records that appeared in MEDLINE (Ovid) from January 1995 to November 2011. The choice of this time period reflected the introduction in the mid-1990s of DW MR sequences into clinical practice. The MEDLINE search strategy included both subject headings (MeSH terms) and text words for the target condition (e.g. stroke, TIA, minor stroke) and the imaging modality under investigation (MRI, DWI). We did not apply any language restrictions. We adapted the MEDLINE search to search EMBASE. In particular, we ‘translated’ the MEDLINE MeSH terms into the corresponding terms available in the Emtree vocabulary. The searches were initially run in November 2010 and updated in November 2011. Full details of both the MEDLINE and the EMBASE search strategies are presented in Appendix 1. We imported all citations identified by the MEDLINE and EMBASE search strategies into the Reference Manager bibliographic database. We hand-searched all proceedings of the International Stroke Conference (2011) and the European Stroke Conference (2011, 2012). We also contacted experts in the field and perused the reference lists of all relevant articles, as well as the most recent issues of Stroke and Cerebrovascular Diseases (not yet indexed in MEDLINE or EMBASE) to identify further published studies for possible inclusion in the review.

Inclusion/exclusion criteria

One review author (MB) examined the titles and abstracts identified by the literature search and retrieved all potentially relevant citations in full. Full-text articles were retained if they focused on primary studies of TIA or minor stroke patients who had DWI as brain imaging in the acute and subacute phase. Studies that did not assess patients by means of DWI, did not report the proportion of patients with positive DWI lesions, or did not report original data, were excluded. In cases of multiple or subsequent publications from the same patient cohort, the most recent or more complete report was chosen.

Two review authors (MB, HM) independently conducted data extraction and reviewed the methodological quality of selected studies. Disagreements were resolved by discussion or referred to a third author (JW). We recorded data on study methods (e.g. setting, study design), characteristics of patients (TIA vs. minor stroke) and imaging findings (DWI-positive brain lesions) as the proportion of patients with an ischaemic lesion on DWI. We also recorded any analyses comparing the prediction of recurrent stroke using the ABCD2 score without and with DWI findings. We also collected information on the following methodological aspects: prospective compared with retrospective study design, definition of TIA, timing of imaging assessment and evaluating clinicians.

Data synthesis

For each study we calculated the total number of patients with a visible DWI lesion. The pooled proportion of TIA patients with a positive DWI lesion was calculated by means of a univariate random-effects meta-analysis with within-study variance modelled as binomial. Heterogeneity between studies was assessed visually by inspection of forest plots and by calculating I²-statistics. 109 We used the statistical software R for our analyses.

Number of included/excluded studies

Diffusion-weighted imaging in patients with transient ischaemic attack/minor stroke

The electronic searches identified 7983 citations. Of these, we considered 185 reports potentially relevant and retrieved the full-text articles for detailed assessment. For two citations, full-text was not easily available and we examined the information contained in the abstracts. Hand-searching of reference lists and of recent issues of relevant medical journal (e.g. European Neurology), not yet indexed in MEDLINE or EMBASE, resulted in a further four reports to be included. We subsequently excluded 127 reports. The most common reason for exclusion was that the study did not provide the number of patients with positive DWI lesions. A total of 45 full-text studies published in 58 reports, and two abstracts, fulfilled our inclusions criteria (Figure 21).

FIGURE 21.

Study identification and selection.

Characteristics of included studies

The methodological characteristics of the included studies are shown in Table 21. Eighteen were observational prospective cohort studies, 17 were observational retrospective cohort studies and 12 studies did not clearly state the type of study design (see Table 21).

All studies used a time-based definition of TIA as opposed to the recently new proposed tissue-based definition. 176

Two studies assessed population-based cohorts, seven studies assessed hospital-based cohorts, eight studies assessed patients from EDs, 27 studies assessed patients from specialist stroke or neurology units and three studies did not clearly provide this information. TIA diagnosis was made by a neurologist in 20 studies, by a stroke specialist in five studies; by an emergency medicine physician in one study and initially by an emergency medicine physician and subsequently confirmed by a stroke specialist in another study; the status of the diagnosing physician was not reported in the remaining 20 studies. Timing of DWI assessment after symptom onset varied across studies (see Table 21). In 18 studies, TIA patients were assessed within 24 hours of symptom onset, in six studies within 48 hours, in 12 studies within 7 days, in three studies within 2 weeks and in the remaining two studies within 3 weeks. Six studies did not clearly provide this information.

Main findings

The total number of TIA/minor stroke patients assessed in the included studies was 9078 (range 18 to 1693 patients). All studies except eight57,72,134,224,225,232,233,248 restricted inclusion to TIA patients. The frequency of positive DWI findings ranged from 9% to 67% across studies (Table 22). Overall, 3048/9078 TIA patients had a visible lesion on DWI. Using a univariate random-effects meta-analysis, the pooled proportion of TIA patients with a positive DWI lesion was 34.3% (95.0% CI 30.5% to 38.4%), with heterogeneity between studies demonstrated by both the forest plot and a high I²-statistic of 89.3%. The forest plot of all included studies is shown in Figure 22.

FIGURE 22.

Proportion of patients with TIA and visible ischaemic lesion on DWI. (Note: Ay and colleagues87 includes Ay and colleagues, 2002,247 2005246).

Two studies also reported findings on CT:138,251 Chatzikonstantinou and colleagues138 found that 2/122 patients had a relevant visible ischaemic lesion on CT (1.6%), compared with 21/122 (17.0%) on DWI; Förster and colleagues251 found that 7/161 (4%) of patients had a visible ischaemic lesion on CT, compared with 71/215 (33%) on DWI. A further study reported only old lesions on CT,93 and one study used CT to exclude non-stroke lesions. 226

Studies varied considerably in terms of sample size, time interval between symptom onset and DWI, symptom duration, aetiology, selection criteria, and setting. We explored potential reasons for heterogeneity of DWI-positive rates between studies including time and inclusion of minor stroke as well as TIA. Four studies provided data separately for TIA and minor stroke patients. 224,225,232,233 The proportion of minor stroke patients with visible ischaemic lesions on DWI was 47% in one study,227 70% in another study,232 and reached 100% in two studies. 224,233 It is worth noting that the two studies in which all minor stroke patients showed a positive DWI lesion were of small sample size and enrolled patients with a definite diagnosis of TIA or minor stroke as opposed to a possible diagnosis of TIA or minor stroke – for example Kastrup and colleagues224 included only patients with high-grade symptomatic carotid stenosis. Similarly, the study by Schulz and colleagues,232 in which 70% of minor stroke patients had a positive DWI lesion, focused on patients with a definite TIA or stroke and scanned patients 3 days after symptom onset or later. In another four studies that included patients with minor stroke as well as TIA but did not report the results separately, the proportion of patients with visible ischaemic lesions on DWI was not noticeably higher than in the studies reporting TIA alone: 53.0%,134 28.0%,57 39.3%72 and 39.0% respectively. 248

We examined the DWI-positive rate by time between symptom onset and scanning (Figure 23) using aggregate data from the primary studies, which showed no obvious time trend although most studies scanned patients within 8 days. Within those 8 days, the proportion with a DWI lesion ranged from 0.09 to 0.68.

FIGURE 23.

Scatterplot of time to imaging vs. proportion of patients with a positive DWI scan. Average time may indicate either a mean or median time.

There is also no readily apparent difference between studies that reported average times and those that reported maximum times to imaging. Attempts to investigate the relationship between proportion DWI positive and time to imaging with statistical modelling did not manage to meet modelling assumptions.

Diffusion-weighted imaging in patients with minor stroke

From the results of our literature search we identified a further four studies that provided data on patients with minor stroke alone, i.e. not TIA. 45,252–254 Three of these studies were performed in the same centre in a regional NV clinic over a total of 12 years (1998–2011)252–254 and included a total of 627 patients (mean 209, range 153–246 patients). The patients were all assessed by expert stroke physicians and neurologists and were discussed at a weekly multidisciplinary meeting, and patients with non-vascular causes of minor stroke (i.e. mimics) were carefully excluded. The patients were scanned on two different MR scanners, about one-third on one scanner and the other two-thirds on the other scanner, but both used the same sequences. The time to scanning varied from within 24 hours to several weeks, but all were scanned on the day of presentation to the clinic. The average proportion of patients with a visible ischaemic lesion on DWI was 69%. Schulz and colleagues50 studied 51 patients with minor stroke and found that 29 patients had an ischaemic lesion on DWI at an average of 3 weeks after the stroke (57%), this lower proportion probably reflecting the greater delay to scanning. Chalela and colleagues32 included some patients with minor stroke [median National Institutes of Health Stroke Scale (NIHSS) score 3] in their prospective comparison of CT and MR DWI, but their assessment of sensitivity and specificity is confounded by switching the diagnosis from TIA to stroke in TIA patients with an ischaemic lesion on DWI.

ABCD2 score and diffusion-weighted imaging or computed tomography finding

Several publications assessed the use of MR DWI findings added to the ABCD2 score (Table 23). We have not re-meta-analysed these studies independently but present a summary of all studies providing information on ABCD2 and DWI findings here. In general, the addition of DWI findings to the ABCD2 score improved the risk prediction (higher AUC values in Table 23). A multicentre study184 assessed the early risk of recurrent stroke after TIA, and performance of ABCD2 score subcategorised as tissue positive or negative on either DWI or CT imaging. Among 3206 patients with DWI, 27.6% (884 patients) had an acute brain infarction. In 1368 patients imaged with CT, 23.9% (327 patients) had an acute or old infarction. For patients with DWI imaging, pooled rates of recurrent stroke at 7 days were 7.1% (95% CI 5.5% to 9.1%) in those with tissue-positive events compared with 0.4% (95% CI 0.2% to 0.7%) in those with tissue-negative events (p < 0.0001). For patients imaged with CT, rates were 12.8% (95% CI 9.3% to 17.4%) and 3.0% (95% CI 2.0% to 4.2%), respectively (p < 0.0001).

There was no information on DWI in TIA/minor stroke mimics, as these patients were excluded from all studies.

Patients with infarction on brain imaging had higher ABCD2 scores than those without any lesion. Tables 24–26 show the risk of recurrent stroke stratified by the presence of brain infarction on imaging and dichotomised ABCD2 score (above and below 4). Data were derived from the systematic review conducted by Giles and colleagues in 2011,210 who reported published and unpublished TIA data from 12 independent centres. Patients with ABCD2 score ≥ 4 and visible lesion on CT or DWI had a 7-day risk of stroke of 10.7%, compared with 1.9% in those without visible lesion; for patients with ABCD2 score of < 4 the risk of stroke was 2.3% and 0.2%, respectively. For patients with visible infarction on DWI and ABCD2 score ≥ 4, the 7-day risk of stroke was 8.9% compared with 0.6% in those without visible lesion; in patients with an ABCD2 score of < 4 these risks were 1.8% and 0.1%, respectively. For patients with visible infarction on CT and ABCD2 score ≥ 4, the 7-day risk of stroke was 15.5%, compared with 4.2% in those without visible lesion, and 3.9% and 0.9% in patients with ABCD2 score of < 4, respectively. The presence of infarction on DWI or CT imaging was the main determinant of early recurrent stroke; however, ABCD2 score had predictive value in the acute phase in both tissue-positive and tissue-negative patients, including both MRI and CT imaging, with an AUC for the prediction of stroke at 7 days of 0.68 (95% CI 0.63 to 0.73) and 0.73 (95% CI 0.67 to 0.80), respectively (p < 0.0001 for both results); AUCs for prediction of recurrent stroke at 90 days were 0.66 (95% CI 0.61 to 0.71) and 0.69 (95% CI 0.62 to 0.76), respectively. 210

Objectives

We performed this systematic review in accordance with Meta-analysis Of Observational Studies in Epidemiology (MOOSE)116 and PRISMA264 statements. We aimed to identify all studies reporting:

• the proportion of patients with transient or mild non-disabling neurological symptoms with a final diagnosis of a non-vascular cause of their symptoms, irrespective of the study design or setting

• the main causes (i.e. differential diagnosis) of non-CV events

• the short and long-term prognosis of non-CV events.

Identification of studies

We identified potential relevant studies by searching Ovid MEDLINE and EMBASE from 1980 to December 2011 (see Appendix 1). We included the following search terms: transient ischaemic attack, TIA, atypical TIA, non-focal TIA, TIA mimics, transient neurological attack or non-CV. Considering that the number of studies assessing the prognosis of patients with non-CV events could be small, first we used a search strategy with broad definitions of TIA, minor stroke and non-CV events in order to retrieve the highest number of articles, and subsequently we included specific search terms for prognosis in these patients. We also screened the reference lists of all relevant studies and review articles. We excluded non-English articles for which a full-text translation was not available.

In the expectation that the literature might be biased towards highly specialist practice that was not relevant to routine clinics, we also identified several sources of unpublished data from operational clinics. These included a prospective data set of patients referred to a local stroke prevention clinic (WGH, Edinburgh, UK) from July 2009 to June 2011, who received a final diagnosis of TIA or minor stroke. 173 This data set is a subset of the SSCAS initiative that has collected anonymised information to monitor stroke care from all hospitals managing acute stroke in Scotland since 2002. We also identified data from an audit undertaken in Leicester.

Inclusion/exclusion criteria

One reviewer (HM) screened the identified titles and retrieved all potentially relevant citations in full-text. Articles were included in this systematic review if they provided any data on proportions of patients diagnosed as non-CV, TIA/minor stroke and also if those studies included a control group of healthy subjects or patients with an alternative condition as long as these subject categories could be clearly distinguished. Studies that did not include patients with non-CV events, did not report the proportion of patients with this condition, or published in abstract form were excluded. In cases when multiple publications were available for the same patient cohort, the most complete report was selected. Duplicates were removed using Reference Manager.

Quality assessment and data extraction

From the selected studies, one reviewer (HM) sought and extracted data on the following variables:

• sample size of study

• study design

• setting of diagnosis

• whether the final diagnosis had been made by a neurologist, geriatrician, stroke physician, GP or nurse

• use of predefined diagnostic criteria for TIA or stroke and method of final diagnosis (i.e. clinical records, in person)

• proportion of patients with a final diagnosis of a non-CV event

• most frequent differential diagnoses in patients with non-CV events

• prognosis of patients with non-CV events, in particular risk of further stroke, MI, and vascular death

• method of follow-up (i.e. telephone, records, in person).

We contacted authors of one study for further information about the person making diagnosis of TIA or non-CV event. For the purpose of this review, patients with diagnosis of ‘possible TIA’ were included in the TIA group, as, in our experience, these patients are treated as TIA in clinical practice. Series including either TIA or minor stroke patients were considered for this review, in light of the frequent description of both conditions in studies reporting the risk of vascular events (i.e. recurrent stroke) after TIA, especially in the context of outpatient clinics.

Data synthesis

We aimed to investigate whether risk of outcome (stroke or vascular death at 7, 90 and > 90 days after presentation) varied between TIA, non-CV and asymptomatic/control patients. Specifically, we compared outcomes between TIA and non-CV patients, and also between non-CV and asymptomatic/control patients on which data on the above outcomes were available. We used simple logistic regression, as the outcomes were binary and the data were not amenable to more complex analyses, to allow for within-study clustering. We also assessed heterogeneity between studies owing to different settings (TIA clinics, hospital clinics and population-based studies) visually with forest plots and calculated I²-statistics. 109 Data were analysed in SAS and R.

Number of included/excluded studies

The literature searches identified 1775 citations. After initial screening of titles and abstracts we selected 53 citations for full-text retrieval. An additional six studies were included after hand-searching of reference lists of relevant studies, yielding a total of 59 studies. After full-text examination, 43 studies were excluded, leaving 16 studies that fulfilled the inclusion criteria. The reasons for exclusion were: not relevant (n = 10), non-CV events excluded (n = 13), review articles (n = 5), conference abstracts (n = 13), risk of duplicated data (n = 1) and no description of the person making diagnosis of non-CV event (n = 1) (Figure 24). None was excluded because of language.

FIGURE 24.

Study selection and reasons for exclusion.

Characteristics of included studies

The 16 included studies7,73,122,156,161,173,257,263,265–271 varied in terms of their primary objectives. Three were population-based studies assessing the incidence of TIA in the general population in UK, the Netherlands and Spain (OCSP, Rotterdam Study and a community-based register in Segovia, respectively). 7,263,265 Eight studies focused on patients attending TIA clinics,122,173,257,266–270 and one study183 enrolled patients from multiple settings as part of a population-based study, even although the final diagnosis was made in a TIA clinic (North Dublin TIA Study). Two were ED-based studies,156,161 one included patients from ED and TIA clinics and reported the results for each setting separately and so contributed two independent data points,73 and one recruited patients as part of a hospital-based clinical trial. 271

The 16 studies included a total of 14,542 participants, and all provided data on frequency of mimics compared with true TIA/minor stroke;7,73,122,156,161,173,183,257,263,265–271 11 studies with a total of 3786 participants reported the main causes of non-CV events7,122,156,161,173,263,266–270 and five studies with a total of 8596 participants reported prognosis data. 122,156,265,270,271 One study deliberately excluded patients with some specific diagnoses other than TIA (i.e. migraine, epilepsy or vertigo). However, other non-specific transient neurological symptoms were incorporated (e.g. positive visual phenomena, dizziness or amnesia) and, consequently, the study was included in this review. 265 The same study used the term ‘mixed TIA’ for patients with focal and non-focal brain symptoms. For the statistical analysis, these patients were included in the group with focal brain symptoms (n = 38, 0.6% of the study sample). Forty-one per cent of the data related to frequency and 66% of the data related to differential diagnosis came from two studies,173,270 and 41% of the data for prognosis came from one study. 271 Follow-up was at least 1 year in all but one study assessing prognosis.

The study design was prospective in 13 studies and retrospective in three (Table 27). The final diagnosis of TIA or non-CV event was made by a neurologist, geriatrician or stroke physician in 15 studies, and one study reported the results of an anterior circulation TIA clinic led by a specialist NV nurse. 268 The final diagnosis of true compared with non-CV event was done in person in 14 studies (from consecutive referrals in eight studies); in two of them a predefined symptoms checklist was used [derived from the Face Arm Speech Test (FAST) in one study]. In two cohorts the final diagnosis was made using clinical records. Thirteen studies used time-based criteria of TIA, whereas three studies did not report which criteria were used. Subjects with minor stroke were included in 13 studies; 10 described the proportion of these events in all true CV events, ranging from 6% to 72%.

Main findings

Frequency of non-cardiovascular events

Figure 25 shows the proportion of patients with a non-CV event in 17 cohorts reported in 16 studies (range 18% to 55%). The summary estimate for proportion of non-CV events was 34% (95% CI 28% to 42%); however, there was evidence of substantial heterogeneity across all selected studies – the I²-statistic was very high, 98.4%. The lowest proportion of non-CV events (14%) was described in a TIA clinic led by a specialist nurse and the highest (55%) in an ED-based study. The proportion of non-CV events was 40% (95% CI 23% to 60%) for TIA clinics, 40% (95% CI 23% to 60%) for population-based studies and 32% (95% CI 10% to 66%) for hospital-/ED-based studies, but with evidence of significant heterogeneity between studies in each setting (Figures 26–28 and Table 28) – the I²-statistics were 97.8%, 95.2%, and 97.9%, respectively, for hospital-/ED-based studies, population-based studies and TIA clinics.

FIGURE 25.

Proportion of patients with a non-CV diagnosis reported in the literature. NA, not applicable.

FIGURE 26.

Hospital/ED-based studies and proportion of non-CV events.

FIGURE 27.

Population-based studies and proportion of non-CV events.

FIGURE 28.

Transient ischaemic attack clinic studies and proportion of non-CV events. NA, not applicable.

TABLE 28 - Proportion of non-CV events in different settings

Study location	Non-CV events (n)	Total patients (n)	Proportion of non-CVs
			(95% CI)
Hospital – ED
Koudstaal and colleagues, 1992271	572	3127	0.18 (0.17 to 0.19)
Ferro and colleagues, 199673	17	31	0.55 (0.38 to 0.71)
Amort and colleagues, 2011156	55	303	0.18 (0.14 to 0.23)
Olivot and colleagues, 2011161	108	224	0.48 (0.42 to 0.55)
Summary			0.32 (0.10 to 0.66)
Population based
Dennis and colleagues, 19897	271	512	0.53 (0.49 to 0.57)
Sempere and colleagues, 1996263	78	313	0.25 (0.21 to 0.30)
Bos and colleagues, 2007262	228	548	0.42 (0.38 to 0.46)
Sheehan and colleagues, 2009183	257	594	0.43 (0.39 to 0.47)
Summary			0.40 (0.23 to 0.60)
TIA clinics
Ferro and colleagues, 199673	16	52	0.31 (0.20 to 0.45)
Martin and colleagues, 1997257	135	508	0.27 (0.23 to 0.31)
Karunaratne and colleagues, 1999266	64	128	0.50 (0.41 to 0.59)
Murray and colleagues, 2007267	395	811	0.49 (0.46 to 0.52)
Lavallée and colleagues, 2007122	240	1085	0.22 (0.20 to 0.25)
Banerjee and colleagues, 2009268	40	282	0.14 (0.10 to 0.19)
WGH 2011	841	2368	0.36 (0.34 to 0.38)
Hörer 2011269	54	123	0.44 (0.36 to 0.53)
Cameron and colleagues, 2011270	1643	3533	0.47 (0.45 to 0.49)
Summary			0.40 (0.23 to 0.60)