Clinically oriented dual-tier screening for post-stroke epilepsy with interpretable machine learning in a severely imbalanced cohort

Background Post-stroke epilepsy is an important complication after stroke, yet its relatively low incidence makes early identification difficult in routine practice. This challenge is compounded by severe class imbalance, which may limit the clinical usefulness of conventional prediction models. Methods We conducted a retrospective cohort study of 21,459 patients with stroke, including 936 patients who developed post-stroke epilepsy. Candidate predictors underwent staged feature reduction, after which we developed an interpretable machine-learning framework using an imbalance-aware modeling strategy. Two clinically distinct models were defined within a dual-tier screening framework: a primary model for balanced risk stratification and a secondary model for sensitivity-prioritized alerting. Model interpretation was examined using SHAP. Results The cohort showed marked class imbalance (approximately 21.9:1). The primary model achieved the most balanced overall performance, with a macro-area under the curve of 0.996, area under the precision-recall curve of 0.970, F1-score of 0.931, sensitivity of 0.907, and specificity of 0.998. The secondary alert model yielded higher sensitivity (0.971) with lower F1-score (0.854) and specificity (0.985), supporting its role as a high-sensitivity screening tool rather than a general prediction model. Key contributors included neurological severity, hypertension, lactate, D-dimer, and aspartate aminotransferase. Conclusion In a severely imbalanced cohort, a clinically oriented dual-tier framework provided both balanced risk stratification and high-sensitivity alerting for post-stroke epilepsy. This approach may support decision-making in follow-up care, although external validation remains necessary.