A group of Stanford Drugs researchers have launched SleepFM Scientific, a multimodal sleep basis mannequin that learns from medical polysomnography and predicts long run illness danger from a single evening of sleep. The analysis work is printed in Nature Drugs and the group has launched the medical code because the open supply sleepfm-clinical repository on GitHub underneath the MIT license.
From in a single day polysomnography to a normal illustration
Polysomnography information mind exercise, eye actions, coronary heart indicators, muscle tone, respiration effort and oxygen saturation throughout a full evening in a sleep lab. It’s the gold normal check in sleep medication, however most medical workflows use it just for sleep staging and sleep apnea analysis. The analysis group deal with these multichannel indicators as a dense physiological time collection and prepare a basis mannequin to be taught a shared illustration throughout all modalities.
SleepFM is educated on about 585,000 hours of sleep recordings from about 65,000 folks, drawn from a number of cohorts. The most important cohort comes from the Stanford Sleep Drugs Middle, the place about 35,000 adults and youngsters had in a single day research between 1999 and 2024. That medical cohort is linked to digital well being information, which later allows survival evaluation for lots of of illness classes.
Mannequin structure and pretraining goal
On the modeling degree, SleepFM makes use of a convolutional spine to extract native options from every channel, adopted by consideration primarily based aggregation throughout channels and a temporal transformer that operates over brief segments of the evening. The identical core structure already appeared in earlier work on SleepFM for sleep staging and sleep disordered respiration detection, the place it confirmed that studying joint embeddings throughout mind exercise, electrocardiography and respiratory indicators improves downstream efficiency.
The pretraining goal is depart one out contrastive studying. For every brief time section, the mannequin builds separate embeddings for every modality group, similar to mind indicators, coronary heart indicators and respiratory indicators, after which learns to align these modality embeddings in order that any subset predicts the joint illustration of the remaining modalities. This method makes the mannequin sturdy to lacking channels and heterogeneous recording montages, that are frequent in actual world sleep labs.
After pretraining on unlabeled polysomnography, the spine is frozen and small job particular heads are educated. For normal sleep duties, a light-weight recurrent or linear head maps embeddings to sleep levels or apnea labels. For medical danger prediction, the mannequin aggregates the complete evening right into a single affected person degree embedding, concatenates fundamental demographics similar to age and intercourse, after which feeds this illustration right into a Cox proportional hazards layer for time to occasion modeling.
Benchmarks on sleep staging and apnea
Earlier than transferring to illness prediction, the analysis group verified that SleepFM competes with specialist fashions on normal sleep evaluation duties. Prior work already confirmed {that a} easy classifier on prime of SleepFM embeddings outperforms finish to finish convolutional networks for sleep stage classification and for detection of sleep disordered respiration, with features in macro AUROC and AUPRC on a number of public datasets.
Within the medical research, the identical pretrained spine is reused for sleep staging and apnea severity classification throughout multi heart cohorts. Outcomes reported within the analysis paper present that SleepFM matches or exceeds present instruments similar to conventional convolutional fashions and different automated sleep staging methods, which validates that the illustration captures core sleep physiology and never solely statistical artifacts from a single dataset.
Predicting 130 illnesses and mortality from one evening of sleep
The core contribution of this Stanford’s analysis paper is illness prediction. The analysis group maps analysis codes within the Stanford digital well being information to phecodes and defines greater than 1,000 candidate illness groupings. For every phecode, they compute time to first analysis after the sleep research and match a Cox mannequin on prime of SleepFM embeddings.
SleepFM identifies 130 illness outcomes whose dangers are predictable from a single evening of polysomnography with sturdy discrimination. These embrace all trigger mortality, dementia, myocardial infarction, coronary heart failure, persistent kidney illness, stroke, atrial fibrillation, a number of cancers and a number of psychiatric and metabolic problems. For a lot of of those situations, efficiency metrics similar to concordance index and space underneath the receiver working curve are in ranges corresponding to established danger scores, despite the fact that the mannequin makes use of solely sleep recordings plus fundamental demographics.
The reporting additionally notes that for some cancers, being pregnant problems, circulatory situations and psychological well being problems, predictions primarily based on SleepFM attain accuracy ranges round 80 p.c for multi yr danger home windows. This means that delicate patterns within the coordination between mind, coronary heart and respiration indicators carry details about latent illness processes that aren’t but clinically seen.
Comparability with easier baselines
To evaluate added worth, the analysis group in contrast SleepFM primarily based danger fashions with two baselines. The primary makes use of solely demographic options similar to age, intercourse and physique mass index. The second trains an finish to finish mannequin straight on polysomnography and outcomes, with out unsupervised pretraining. Throughout most illness classes, the pretrained SleepFM illustration mixed with a easy survival head yields increased concordance and better lengthy horizon AUROC than each baselines.
This analysis clearly exhibits that the achieve comes much less from a posh prediction head and extra from the inspiration mannequin that has discovered a normal illustration of sleep physiology. In apply, which means medical facilities can reuse a single pretrained spine, be taught small web site particular heads with comparatively modest labeled cohorts and nonetheless method state-of-the-art efficiency.
Try the Paper and FULL CODES right here. Additionally, be at liberty to observe us on Twitter and don’t overlook to affix our 100k+ ML SubReddit and Subscribe to our Publication. Wait! are you on telegram? now you possibly can be a part of us on telegram as properly.
Try our newest launch of ai2025.dev, a 2025-focused analytics platform that turns mannequin launches, benchmarks, and ecosystem exercise right into a structured dataset you possibly can filter, examine, and export
Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is dedicated to harnessing the potential of Synthetic Intelligence for social good. His most up-to-date endeavor is the launch of an Synthetic Intelligence Media Platform, Marktechpost, which stands out for its in-depth protection of machine studying and deep studying information that’s each technically sound and simply comprehensible by a large viewers. The platform boasts of over 2 million month-to-month views, illustrating its recognition amongst audiences.
