Cardio Fusion: A Dual-Stream Regularized Fusion Architecture for Generalizable Cardiovascular Disease Prediction Under Dataset Heterogeneity

Computationalrisk-predictionsystems havebecomeanindispensableadjuncttoconventionalclinicaldiagnosis, yet the field continues to be constrained by a persistent methodological obstacle: benchmark datasets remain small, heterogeneous in feature composition, and frequently outdated, while single-classifier models exhibit limited generalizability across diverse patient populations. To address this limitation, the present study proposes a dual-branch hybriddeeplearningarchitecture comprising(i) a tabularbranch,inwhichanL1/L2-regularizeddeepneuralnetworkis ensembled with Random Forest, Support Vector Machine, and K-Nearest Neighbour classifiers over a feature subset optimized through Neighbourhood Component Analysis [13], and (ii) an imaging/ECG branch constructed upon ResNet-50 [14], DenseNet-121 [15], and EfficientNet-B0 [16], with the outputs of both branches integrated via a weighted soft-voting meta-learner. The proposed framework is systematically benchmarked across six contemporary datasets, spanning the canonical Cleveland and Statlog repositories, the 920-instance four-hospital combined UCI dataset that has gained prominence in recent (2024–2025) federated learning and interpretability research, a merged 1,190-recordUCI–Kaggle cohort, the feature-rich Z-AlizadehSanicohort[18], andthelongitudinalFraminghamHeart Studycohort [19]. OnthestandardClevelandbenchmark, the proposedregularizedhybridclassifier attains anaccuracy of 96.5%; on a second, independently re-implemented empirical dataset — the 68,612-record Kaggle cardiovascular cohort (cardio_train.csv) — the fused model achieves an accuracy of 73.95% and an AUC-ROC of 80.57% (95% CI: 79.96–81.14), aresult attributabletothe dataset's exclusively lifestyle- and vitals-basedfeature compositionrather than to any deficiency in the proposed architecture. Across the broader multi-dataset benchmark, the framework demonstrates competitive performance relative to recent state-of-the-art systems, including stacked TabNet–XGBoost ensembles (95.20% accuracy, 91.92% F1-score) [2] andappliedartificial-intelligence pipelines reportingaccuracies of up to 99.48% on the Statlog cohort [4]. Collectively, these findings substantiate the premise that the integration of optimized feature selection, architectural hybridization, and explicit regularization yields a classifier that isconcurrently more accurate and more robust across heterogeneous clinical data sources than conventional single-algorithm baselines, thereby reinforcing its viability as a deployable clinical decision-support tool.