Enhanced Stroke Risk Prediction Through Optuna-Based Hyperparameter Optimization and Ensemble Classification Models

Abstract

Stroke remains a major cause of global mortality and prolonged disability, highlighting the importance of accurate and early risk prediction. This study actively shows a robust and interpretable machine learning framework for stroke classification by incorporating Optuna-based hyperparameter tuning to enhance predictive performance. Structured clinical data were pre-processed through missing-value imputation, categorical encoding, feature normalization, and class imbalance correction using SMOTEENN. Baseline model benchmarking with Lazy Classifier identified Logistic Regression, Ridge Classifier, and Calibrated Logistic Regression as the most promising candidates for optimization. These models were then fine-tuned using Optuna, which efficiently explores the hyperparameter search space while reducing the computational burden associated with traditional grid search. Optimized models were combined using soft voting and stacking ensembles to further improve classification stability and generalization. Performance was evaluated using accuracy, precision, recall, specificity, sensitivity, F1-score, and ROC-AUC. Results show that Optuna-based hyperparameter tuning significantly enhances model robustness and generalization when compared to conventional Grid Search. Among all approaches, the Optuna-tuned stacking ensemble outperformed individual classifiers and soft voting, establishing it as the most reliable, interpretable, scalable, and data-driven solution for clinical stroke risk prediction.