Validation and comparison of imaging-based scores for prediction of early stroke risk after transient ischaemic attack: a pooled analysis of individual-patient data from cohort studies

BACKGROUND: Identification of patients at highest risk of early stroke after transient ischaemic attack has been improved with imaging based scores. We aimed to compare the validity and prognostic utility of imaging-based stroke risk scores in patients after transient ischaemic attack. METHODS: We did a pooled analysis of published and unpublished individual-patient data from 16 cohort studies of transient ischaemic attack done in Asia, Europe, and the USA, with early brain and vascular imaging and follow up. All patients were assessed by stroke specialists in hospital settings as inpatients, in emergency departments, or in transient ischaemic attack clinics. Inclusion criteria were stroke-specialist confirmed transient ischaemic attack, age of 18 years or older, and MRI done within 7 days of index transient ischaemic attack and before stroke recurrence. Multivariable logistic regression was done to analyse the predictive utility of abnormal diffusion-weighted MRI, carotid stenosis, and transient ischaemic attack within 1 week of index transient ischaemic attack (dual transient ischaemic attack) after adjusting for ABCD2 score. We compared the prognostic utility of the ABCD2, ABCD2-I, and ABCD3-I scores using discrimination, calibration, and risk reclassification. FINDINGS: In 2176 patients from 16 cohort studies done between 2005 and 2015, after adjusting for ABCD2 score, positive diffusion-weighted imaging (odds ratio [OR] 3·8, 95% CI 2·1-7·0), dual transient ischaemic attack (OR 3·3, 95% CI 1·8-5·8), and ipsilateral carotid stenosis (OR 4·7, 95% CI 2·6-8·6) were associated with 7 day stroke after index transient ischaemic attack (p<0·001 for all). 7 day stroke risk increased with increasing ABCD2-I and ABCD3-I scores (both p<0·001). Discrimination to identify early stroke risk was improved for ABCD2-I versus ABCD2 (2 day c statistic 0·74 vs 0·64; p=0·006). However, discrimination was further improved by ABCD3-I compared with ABCD2 (2 day c statistic 0·84 vs 0·64; p<0·001) and ABCD2-I (c statistic 0·84 vs 0·74; p