Adaptive sampling enhanced deep learning framework for accurate interpretable stroke risk prediction

Stroke is a leading cause of global mortality and long-term disability, emphasizing the urgent need for predictive models that are accurate, interpretable, and equitable to support precision medicine. Conventional risk assessment methods often rely on a limited set of clinical indicators and ignore subgroup-specific patterns, which reduces predictive performance and can bias outcomes against underrepresented populations. To address these challenges, this study proposes ASTab-Stroke (Adaptive Stratified TabNet for Stroke Prediction), a deep learning framework integrating Adaptive Stratified Sampling (ASS) with TabNet’s sequential attention mechanism. ASS dynamically reweights patient strata based on their contribution to prediction errors, ensuring fair representation of minority and high-risk subgroups without introducing synthetic data. TabNet’s sequential attention provides step-wise feature attribution, enabling clinicians to interpret the influence of predictors such as age, hypertension, heart disease, glucose level, BMI, and lifestyle factors on stroke risk. The framework was implemented in Python 3.10 and evaluated using the Stroke Prediction Dataset, which includes diverse demographic, clinical, and lifestyle variables. ASTab-Stroke achieved 98% accuracy, 0.998 AUC, 0.97 F1-score, 0.99 recall, and 0.98 precision, outperforming existing baselines by approximately 3% in accuracy while demonstrating improved sensitivity and fairness across clinically significant subgroups. The age and comorbidity features proved to be critical in ablation studies and work on cross-validation showed strong generalization. This framework is a clinically interpretable, scalable, and ethically rationalized method of stroke risk prediction, which gives dependable information to support clinical decision-making with data. The flexibility of it implies that it has a wide potential to be used in other fields of precision medicine, where interpretability and subgroup fairness are crucial in promoting equitable and informed patient care.