Multi-label auxiliary decision-making model for pharmacotherapy of ischemic stroke based on BR-SA-LightGBM using real-world electronic medical records

Background Ischemic stroke poses a substantial global health burden owing to its high prevalence, recurrence, disability, and mortality rates. Pharmacotherapy for ischemic stroke is highly heterogeneous across patients, with complex polypharmacy regimens and intricate interdependencies between drug classes. Precise, personalised medication recommendations are therefore indispensable for optimising therapeutic outcomes and improving patient prognosis. Methods This study framed pharmacotherapeutic decision-making for ischemic stroke as a multi-label learning task. Using retrospective electronic medical records from 7,335 ischemic stroke patients admitted to a single tertiary hospital (Cangzhou Central Hospital), we developed and evaluated an integrated model, termed BR-SA-LightGBM, which combines Binary Relevance (BR), Simulated Annealing (SA), and the Light Gradient Boosting Machine (LightGBM) to provide personalised recommendations across eight first-line drug categories. Results In an independent test cohort ( n = 2,200), the BR-SA-LightGBM model achieved an accuracy of 0.90, an F1-score of 0.87, and a markedly low Hamming loss of 0.09. It significantly outperformed ML-KNN, Power-Set, and the unoptimized BR-LightGBM model across all evaluation metrics ( P < 0.05). Across drug classes, age, the number of comorbidities, atrial fibrillation, carotid plaque, and fibrinogen were among the most influential predictors of medication selection. Conclusions The BR-SA-LightGBM model established herein enables accurate, personalised, auxiliary decision support for drug prescribing in ischemic stroke. It is intended as an auxiliary decision-support tool that has the potential to reduce medication errors and omissions and may assist clinicians in improving clinical efficiency; it also provides a novel methodological framework for precision medicine in stroke care.