ABSTRACT NUMBER: ESOC2026LT67 ARTIFICIAL INTELLIGENCE BASED FUTURE STROKE RISK PREDICTION USING ROUTINELY COLLECTED ECHOCARDIOGRAM REPORTS

Abstract Background and aims Stroke is a leading cause of death and disability in the UK. Despite preventive medication and lifestyle modification reducing stroke risk, approximately 30% of strokes remain unexplained. Therefore, better identification of individuals at risk of stroke is needed. This presentation reports preliminary phase 1 results from A.I. BASED STROKE RISK FACTOR CLASSIFICATION AND TREATMENT (ABSTRACT), UK Medical Research Council (MRC) funded study. ABSTRACT focuses on using Artificial Intelligence (AI) on a multimodal dataset derived from routinely collected hospital data for stroke risk prediction. Here, we focus on single modality analyses using Echocardiogram (Echo) reports, with models trained independently on three data sources prior to dataset merging. Methods Historical stroke cases admitted in southwest England between January 2003 - November 2025 were identified using the Sentinel Stroke National Audit Programme (SSNAP) database. Data analysis was performed for 9,155 stroke patients and 109,581 controls. Echo reports were obtained from three sources: University Hospitals Plymouth (UHP), Phillips and Ultracardiac. For this interim analysis, eXtreme Gradient Boosting (XGBoost) models were trained on datasets individually for 1 year stroke risk prediction. Cross-dataset harmonizations and integrated modelling are currently ongoing. Results Provisionally, across the datasets, models demonstrated prediction accuracy within 0.6-0.71 range for 1 year stroke risk prediction. Preliminary interpretability analysis using SHapley Additive exPlanations (SHAP), suggested that left atrial structure and left ventricular diastolic function were influential across all datasets. Conclusions This phase 1 analysis provides early promising findings across datasets, supporting ongoing harmonisation and future integrated, multimodal modelling within this large study. Conflict of interest