Tabular Feature Guided Multimodal Cardiovascular Disease Classification

Cardiovascular diseases (CVD) remain a critical global health challenge, necessitating innovative approaches for early screening and prevention. Sleep, a complex physiological process intrinsically linked to cardiovascular and cerebrovascular systems, offers a promising avenue for health monitoring through emerging wearable technologies and multimodal data integration. Despite significant advancements in machine learning and deep learning methodologies, existing CVD prediction approaches predominantly rely on single-modality data, thereby limiting comprehensive health assessment. Traditional methods often struggle to effectively integrate heterogeneous data sources, including high-resolution physiological signals, demographic information, and manually derived features. To address these limitations, we propose a novel tabular feature guided multimodal cardiovascular disease prediction framework that leverages multiple data modalities during sleep monitoring. Our approach innovatively employs multi-granularity encoders and develops a knowledge-driven tabular feature guided fusion module, enabling sophisticated feature extraction and integration across different hierarchical levels. Experimental validation on public datasets demonstrates that the proposed model outperforms existing multimodal models, achieving a 7.29% improvement in accuracy. This underscores its potential for advancing cardiovascular disease prediction and personalized healthcare monitoring.