A Cross-Modal Contrastive Representation Learning Model for OSAHS Detection Based on Decoupling and Reconstruction Operations

Obstructive Sleep Apnea-Hypopnea Syndrome (OSAHS) significantly impacts public health and safety. Multimodal biomedical signals (MBS) are essential for OSAHS screening. However, current deep learning frameworks for MBS struggle to differentiate between modality-specific and shared information, leading to confusion and reduced model performance. Additionally, the heterogeneity of modalities and limited labeled data further hinder effective MBS modeling. To address these challenges, we propose a cross-modal contrastive learning method based on decoupling and reconstructing multimodal biomedical signals (CMDR). Specific encoders extract unique information from different modalities, while specific decoders handle modality reconstruction. A shared encoder then extracts patterns across modalities, using a cross-attention mechanism to decouple relationships and ensure orthogonality between modality-specific and shared features. Cross-modal contrastive learning is applied to the representations from each specific encoder, ensuring effective extraction and learning of modality-specific information. Finally, specific and shared features are fused, with a self-attention mechanism capturing both specificity and consistency of the signals. We validate CMDR on three multimodal biomedical signal datasets for OSAHS screening. Experimental results demonstrate that CMDR outperforms current baseline models, achieving an average F1-score improvement of 9.27%, setting a new state-of-the-art in OSAHS task performance.