Interpretable Cardiovascular Diagnosis using Multi-dimensional Feature Fusion and Deep Learning

Abstract

This research aims to improve cardiovascular disease diagnostic accuracy and interpretability by developing a multi-dimensional feature fusion model that captures the complex, multi-faceted nature of cardiovascular conditions. The framework integrates five modalities — Electrocardiogram (ECG), Photoplethysmography (PPG), echocardiogram video, heart sounds, and clinical text—using modality-specific neural networks for feature extraction. These features are consolidated via feature-level concatenation and processed through a Multi-Layer Perceptron (MLP) classifier. SHapley Additive exPlanations (SHAP) analysis was subsequently employed to evaluate individual modality contributions and ensure clinical transparency. Testing against public databases demonstrated a peak diagnostic accuracy of 96.8%. This performance significantly outperformed all unimodal and partial-modal benchmarks across key performance metrics, including precision, recall, and F1-score. To provide clinical interpretability, SHAP analysis was utilized to quantify the contribution of each modality to the final prediction. The analysis revealed that echocardiogram and ECG data provided the highest predictive power within the multi-modal framework. By successfully consolidating disparate biomedical signals, this approach provides a robust path for advanced diagnostics. Future development will focus on privacy-preserving architectures and the integration of these models into wearable technology for real-time, remote patient monitoring systems, ensuring the model remains viable for clinical environments. This framework uniquely integrates five distinct biomedical modalities with SHAP interpretability, establishing a revolutionary diagnostic path that outperforms traditional unimodal systems in both accuracy and clinical transparency.