Robust Cross-Domain Multimodal Deep Learning for Biomedical Signal Interpretation and Disease Prediction

Abstract

Biomedical signal analysis is essential for continuous patient monitoring and AI-assisted clinical decision-making using multimodal physiological data. This study proposes a robust cross-domain multimodal deep learning framework that integrates modality-specific encoders, adaptive attention-based fusion, and domain-adversarial learning to improve generalization across heterogeneous biomedical datasets. The framework jointly analyzes electrocardiogram (ECG), photoplethysmography (PPG), and intensive care unit (ICU) physiological signals, including heart rate, blood pressure, respiratory rate, and oxygen saturation, from the publicly available PTB-XL, BIDMC, and MIMIC-IV datasets. Explainability is achieved through Integrated Gradients and modality-level attribution. Experimental results achieved 94.8% accuracy, 93.7% F1-score, and 96.1% ROC-AUC, outperforming CNN, LSTM, Transformer, and tensor fusion baselines. Cross-domain evaluation improved the F1-score by 7–10% over tensor fusion methods while maintaining robustness under dataset shifts, noisy inputs, and missing modalities. These results demonstrate that adaptive multimodal fusion with domain-invariant learning enhances the robustness, interpretability, and clinical applicability of AI-assisted biomedical signal analysis and disease prediction.