MT-ARM-Net: An Explainable Multi-Task Deep Learning Framework for Cardiovascular Risk Assessment using Heterogeneous Dsata

Background and Objective Cardiovascular diseases (CVDs) encompass heterogeneous diagnostic paradigms ranging from acute electrophysiological arrhythmias to chronic clinical risk profiling. While artificial intelligence models achieve strong performance on isolated diagnostic benchmarks, unifying disparate data modalities—specifically high-dimensional electrocardiogram (ECG) time-series and low-dimensional demographic tabular records—within a single neural architecture remains a formidable challenge due to cross-modal gradient competition and class imbalance. Methods This study investigates MT-ARM-Net , an exploratory multi-task deep learning framework integrating convolutional residual networks, multi-head self-attention, and a tabular multilayer perceptron (MLP) via a zero-masked feature routing and self-gating recalibration bottleneck. Formulated as a methodological stress-test for concurrent parameter optimization, the architecture is evaluated across three unlinked public benchmarks: MIT-BIH Arrhythmia (5-class ECG), PTBDB Myocardial Infarction (binary ECG), and Framingham Offspring (binary 10-year Coronary Heart Disease tabular risk). To counteract cross-modal gradient domination by dense ECG streams, a static task-weighting schedule ( ) is enforced alongside a strict leakage-free, patient-stratified train–test partitioning. Results Under a controlled, seed-invariant deterministic evaluation protocol, MT-ARM-Net demonstrates strong discriminative point estimates on electrophysiological tasks, achieving an overall accuracy of (Positive-Class , ) on PTBDB and (Macro , ) on MIT-BIH, performing competitively against standalone knowledge-distilled ( Distill-ARM-Net ) and tree-based ( XGBoost ) baselines. Conversely, on the severely skewed Framingham cohort ( positive prevalence), the unified model exhibits pronounced sensitivity degradation under the default uncalibrated decision threshold ( ), yielding an of and a Positive-Class of ( ). Forensic audit via Local Interpretable Model-agnostic Explanations (LIME) confirms that this performance gap stems from negative cross-modal transfer and protective covariate dominance. Conclusion Ultimately, these findings delineate both the representational viability of shared ECG encoders and the intrinsic boundaries of uncalibrated multi-task learning across skewed tabular domains, emphasizing the necessity of post-hoc probability recalibration for clinical translation.