Multimodal PSG foundation model integrating second to full-night scale

Current sleep deep learning methods are typically limited by their focus on short, isolated segments of data, which overlooks the complex, multifaceted nature of full-night Polysomnography (PSG). To overcome these limitations, we present PSG-M&m, a foundation model designed to analyze both granular, second-by-second signal details and the broader, hour-long structural patterns of sleep. Our model employs a hierarchical dual-encoder architecture: a Macro-Encoder that evaluates long-term temporal trends throughout the night and a micro-Encoder that extracts localized characteristics from biosignals, optimized using a combination of masked autoencoding and multi-modal contrastive learning. Macro-Encoder is refined through Demographic-Guided Contrastive Learning, which enhances its global representation by aligning sleep patterns with patient-demographics. Trained on a vast dataset (>20,000 PSG recordings, 158K hours), PSG-M&m significantly surpasses current foundation models. It offers improved generalizability for downstream clinical tasks, providing a more robust framework for comprehensive sleep analysis.