Fusion-Based Hybrid Meta-Learning: Enhancing Cardiovascular Disease Prediction with AutoML and Explainable AI
Md. Mijanur Rahman, Shahla Farsi, Marguf Khan Tonmoy, Sadia Isfat Ara Rahman, Md Rezaul Karim Khan · 2025
Cardiovascular diseases (CVD) are a consequential widespread health challenge since they have a high incidence and mortality rate. Therefore, early detection and accurate diagnosis are crucial to more effective treatment and better patient interpretability. However, real-world CVD prediction models still lack explainability and transparency. This approach is resolved here through focusing on robustness, accuracy, reliability and human interpretability. A hybrid fusion model was proposed consisting of AutoML tree-based pipeline optimization tools (TPOT), light-grade boost machines (LightGBM), extreme-grade boost machines (XGBoost) and meta-learner to extract multilayer perceptron (MLP) and random forest (RF) to improve predictive heart disease classification. The Cleveland heart disease dataset was employed for training. The Imbalance of data was handled using synthetic minority oversampling technique (SMOTE) and jittering-based augmentation. Polynomial expansion and standardized scaling were used for feature engineering. Soft voting fusion models XGBoost, LightGBM and TPOT were trained with probability calibration (CalibratedClassifierCV) to optimize the confidence scores. The outputs were then fed through a meta-learner MLP and RF to classify. The model achieved 96.75% test accuracy, MCC of 0.9350, an AUC 96.72, and a 10 -fold cross-validation accuracy of 98.67 %, reflecting excellent generalization and robustness. Moreover, Explainable AI (XAI) based on shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME) validated model interpretability and feature ranking. Our work can assist in real-life diagnosis with faster and more precise predictions of CVD using explainable AI inference. This supports early diagnoses and improves confidence and reliability in automated tools within clinics.