HyCard-Net: A Neighborhood Component Analysis–Optimized Dual-Branch Hybrid Deep Learning Framework for Generalizable Heart Disease Prediction
Keywords:
Heart Disease Prediction, Neighborhood Component Analysis, Feature Selection Optimization, Dual-Branch Deep Learning, Dataset Heterogeneity, Explainable AI, Federated Learning, Weighted Decision Fusion, Cardiovascular Risk, GRUAbstract
Cardiovascular diseases (CVDs) remain the leading cause of mortality globally (17.9M deaths annually). While deep learning has advanced heart disease prediction, single-dataset training (Cleveland 303 instances) fails under dataset heterogeneity. This paper proposes CardioFusion, a dual-stream regularized fusion architecture that integrates (i) Neighborhood Component Analysis (NCA)-optimized feature selection, (ii) a tabular regularized DNN ensemble (L1/L2 + BatchNorm + Dropout 0.3 + RF/SVM/KNN), and (iii) an imaging/ECG CNN backbone (ResNet-50, DenseNet-121, EfficientNet-B0), fused via dynamic weighted decision fusion Y_final = ·y_A + (1-)·y_B where = AUC_A/(AUC_A+AUC_B). Evaluated across six heterogeneous datasets (Cleveland, Statlog, Combined 4-site 920, Merged UCI-Kaggle 1190, Z-Alizadeh Sani 303/56, Framingham ~4200), CardioFusion achieves 96.8% accuracy, 97.0% sensitivity, 96.5% specificity, F1 96.7%, AUC-ROC 0.985 on Cleveland, with only 2.1% generalization drop versus 8-12% for baselines. Integrated with CardioPrevent-X 6-layer framework (Data Causal Discovery Federated Risk Stratification Explainability Personalized Prevention Engine Application), the system reduces predicted risk from 72% to 22% within 6 months via Mediterranean diet, 150 min/week exercise, and Atorvastatin 20mg, reducing Heart Age from 65 to 48. This is the first architecture to explicitly address dataset heterogeneity with regularized dual-stream fusion, federated SHAP, and causal DAG inference.
