A Dual-Tier Explainable Meta-Ensemble Framework for Cardiovascular Risk Prediction using Cross-Dataset ECG-Clinical Fusion

Abstract

CVDs are still the first cause of death worldwide, therefore, an early prediction system with high accuracy and interpretability is demanded. Although promising performance has been achieved by machine learning (ML) and deep learning (DL) methods, their translation into clinical applications remains limited due to novelty, heterogeneity across limited generalization across heterogeneous datasets, lack of interpretability, and class imbalance challenges. In this paper, we introduce ExDT-CVRP, a two-tier explainable meta-ensemble model for cardiovascular risk prediction based on cross-dataset fusion of ECG signals and clinical tabular data. The method incorporates pre-processing and statistical feature selection based on ANOVA and chi-square tests, with ADASYN-based class balancing applied during model training. At Stage 2, modality-specific feature representations are extracted using statistical feature engineering for clinical data and a one-dimensional convolutional neural network (1D-CNN) for ECG feature extraction. Cross-dataset representation alignment is then performed, followed by dual-modal feature fusion to construct a unified representation without requiring patient-level correspondence. At Tier-1, a hybrid set of base learners comprising a Deep Neural Network (DNN), Light GBM, Random Forest, XGBoost, and Support Vector Machine (SVM) is employed to learn complementary patterns from the unified dual-modal representation. At Tier-2 a deep neural meta-learner is employed to combine base model predictions to model nonlinear dependencies between the base model predictions and extend robustness and generalization. Model interpretability is pro- vided using SHAP-based explanations. Experimental results on publicly available datasets confirm that the proposed framework outperforms the state-of-the-art with accuracy 96.7%, recall 96.8% and AUC-ROC 98.4%. The key novelty lies in cross-dataset representation alignment, enabling multimodal learning without requiring paired samples. The experimental results support that the presented cross-dataset dual-modal approach can deliver superior prediction results and can be interpretable to a certain extent.