Machine learning in the diagnosis and prognosis of transient ischaemic attack: a systematic review

Transient ischaemic attack (TIA) is a major risk factor for stroke, with up to 15% of patients experiencing an event within 90 days, a large proportion in the first 48 h. Accurate diagnosis and prognostic stratification remain challenging due to transient symptoms, lack of biomarkers, and limitations of traditional clinical scores. Machine learning (ML) holds the potential to enhance diagnostic accuracy, risk prediction, and prognosis by harnessing complex clinical, imaging, and electronic health record data. PubMed was searched from January 2000 to the present. Studies reporting the association between ML models and TIA prognosis and diagnosis were included. Risk of bias was assessed using the PROBAST tool. Data were synthesised narratively due to heterogeneity in outcomes and methodologies. A total of 10 studies were included. ML models demonstrated good discriminatory ability in diagnosing TIA compared with healthy controls. However, performance declined in metabolically complex subgroups and when distinguishing TIAs from stroke mimics. Three studies developed risk prediction models using electronic health records or CTA radiomics, with best-performing models achieving AUCs of 0.82–0.88. Studies that assessed prognostic outcomes consistently outperformed logistic regression, achieving AUCs of 0.77–0.94. ML models show promise in enhancing TIA diagnosis, risk prediction, and prognostication, frequently outperforming traditional clinical scores and statistical methods. However, most studies lacked external validation, used heterogeneous endpoints, and provided limited interpretability. Future work should prioritise larger prospective cohorts, standardised outcome definitions, model explainability, and evaluation of real-world clinical impact. CRD420251123717