Interpretable Incremental Tree-Based Learning for Brain Stroke Prediction with Rule Evolution

Accurate brain stroke prediction requires models that can detect high-risk cases while adapting to newly arriving clinical data. This study proposes an interpretable incremental tree-based learning approach for brain stroke prediction using Decision Tree, Random Forest, and Gradient Boosting. Experiments were conducted using the Cerebral Stroke Dataset, a public imbalanced stroke prediction dataset with 43,400 records and 12 attributes. Since the dataset does not provide verified patient-level timestamps, the data were divided into three sequential simulated cohorts to evaluate model behavior under progressive data availability. Decision Tree was used as a refitted interpretable baseline, while Random Forest and Gradient Boosting were updated using a warm-start strategy by appending additional estimators at later stages. Rule extraction and rule classification were used to translate tree-based decision logic into readable rules and categorize them as conforming, generalized, specialized, unexpected, or novel, enabling analysis of rule evolution across cohorts. Experimental results showed that the incremental approach improved sensitivity oriented performance and reduced update time compared with traditional retraining. At the final stage, incremental Gradient Boosting achieved 98.044% accuracy, 75.32% F1-score, 91.03% recall, and 94.66% ROC AUC, with only 40 false negatives. Therefore, incremental Gradient Boosting provided the most clinically suitable trade-off between stroke-case detection, update efficiency, and interpretable rule behavior.