XSleepFusion: A dual-stage information bottleneck fusion framework for interpretable multimodal sleep analysis

Abstract

Sleep disorders affect hundreds of millions globally, with accurate assessment of sleep apnea (SA) and sleep staging (SS) essential for clinical diagnosis and early intervention. Manual analysis by sleep experts is time-consuming and subject to inter-rater variability. Deep learning (DL) approaches offer automation potential but face fundamental challenges in multi-modal physiological signal integration and interpretability. This paper presents XSleepFusion, a cross-modal fusion framework based on information bottleneck (IB) theory for automated sleep analysis. The framework introduces a dual-stage IB mechanism that systematically processes physiological signals: first eliminating intra-modal redundancy, then optimizing cross-modal feature fusion. An evolutionary attention Transformer network (EAT-Net) backbone extracts temporal features at multiple scales, providing interpretable attention patterns. Experimental validation on eight clinical datasets comprising over 15,000 sleep recordings demonstrates the framework’s effectiveness in polysomnogram (PSG)-based SA detection, electrocariogram (ECG)-based SA detection, and SS. The architecture achieves superior generalization across varying signal qualities and modal combinations, while the dual-stage design enables flexible integration of diverse physiological signals. Through interpretable feature representations and robust cross-modal fusion capabilities, XSleepFusion establishes a reliable and adaptable foundation for clinical sleep monitoring.