Researchers at Stanford University Propose SleepFM: The First Multi-Modal Foundation Model for Sleep Analysis
How does SleepFM improve on traditional CNN models?
SleepFM improves on traditional CNN models by leveraging a large multi-modal dataset and employing a novel leave-one-out contrastive learning approach5. This allows the model to learn robust embeddings and capture the richness of sleep recordings more effectively. As a result, SleepFM outperforms end-to-end CNNs in tasks such as sleep stage classification and sleep-disordered breathing detection.
What is the primary function of SleepFM?
SleepFM is the first multi-modal foundation model for sleep analysis that addresses existing models' limitations. It leverages a large dataset of PSG records from over 14,000 participants to learn robust embeddings through contrastive learning (CL). The model employs a novel leave-one-out approach to CL, which improves the performance of downstream tasks compared to the standard pairwise CL.
What novel approach does SleepFM use in contrastive learning?
SleepFM employs a novel leave-one-out approach in contrastive learning, which improves downstream task performance compared to standard pairwise contrastive learning. This innovative approach allows the model to learn representations by aligning each modality with an aggregate representation of the remaining modalities, encouraging holistic learning of multi-modal data.