Sleep monitoring is a crucial part of maintaining overall health, yet accurately assessing sleep and diagnosing disorders is a complex task due to the need for multi-modal data interpretation typically obtained through polysomnography (PSG). Current methods often depend on extensive manual evaluation by trained technicians, making them time-consuming and susceptible to variability. To address these challenges, researchers from Stanford University and the Technical University of Denmark have developed SleepFM, a model designed to enhance the accuracy of sleep analyses.
Currently, deep learning models using end-to-end convolutional neural networks (CNNs) are the most common method for sleep analysis, but their performance needs improvement when interpreting data from various physiological sources. SleepFM addresses these limitations by leveraging a large PSG dataset from over 14,000 participants and implementing contrastive learning (CL), a method designed to accelerate the learning process of the model.
This new model’s architecture features three CNNs based on EfficientNet architecture, each generating embeddings for different physiological categories: brain activity signals, ECG, and respiratory signals. The novel leave-one-out CL approach aligns each category with an aggregate representation of the remaining categories, encouraging holistic learning of multi-modal data and optimizing the model’s performance in downstream tasks.
When tested, SleepFM significantly outperformed the traditional CNNs. In regards to sleep stage classification, the logistic regression model trained on SleepFM’s embeddings achieved a macro AUROC of 0.88 compared to 0.72 from CNNs, and a macro AUPRC of 0.72 versus 0.48. In detecting sleep-disordered breathing (SDB), SleepFM also achieved higher scores with an AUROC of 0.85 and an AUPRC of 0.77. Additionally, SleepFM excelled in identifying corresponding recording clips from different modalities, demonstrating its ability to effectively interpret multi-modal sleep data.
In conclusion, SleepFM provides an innovative solution to the challenges of sleep monitoring and disorder diagnosis, significantly outperforming traditional CNNs in a variety of sleep-related tasks. With its exceptional performance in sleep stage classification and SDB detection, alongside its effective applicability to external datasets, SleepFM could pioneer novel advancements in sleep research and clinical applications.
The complete research is available on the published paper and on GitHub. Follow them for more updates and consider joining their communities on Telegram, LinkedIn, and Reddit. Lastly, you can contribute towards this research or similar projects by subscribing to their newsletter.
