From unimodal to multimodal deep physiological signal analysis in healthcare: A survey of models, fusion and explainability

Physiological signal analysis supports objective assessment of emotion, stress, pain, and disease progression in healthcare systems. Early research focused mainly on unimodal signals, which perform well in controlled settings but are often affected by noise, limited contextual information and reduced generalisability. Recent advances have shifted towards multimodal deep learning frameworks that jointly exploit complementary physiological signals to improve performance and clinical relevance. This survey provides a comprehensive review of the transition from unimodal to multimodal deep physiological signal analysis in healthcare and systematically examines commonly used physiological modalities, benchmark datasets and evaluation metrics. The survey further analyses multimodal alignment strategies and fusion techniques, covering early, intermediate, and late fusion paradigms, with an application-oriented discussion spanning emotion recognition, stress detection, pain assessment and other related domains. Particular emphasis is placed on explainable artificial intelligence methods for multimodal physiological models, highlighting modality-level attribution, cross-modal interactions and application-specific interpretability. Finally, it discusses key dataset- and method-related limitations, such as data scarcity, class imbalance, fusion complexity and computational constraints and outlines future research directions related to wearable edge intelligence, privacy-aware learning, enhanced data diversity and quantum-inspired approaches. This survey can be used as a structured reference for researchers and practitioners developing trustworthy multimodal deep learning solutions for next-generation healthcare systems.