Trained on 12,392 nights of polysomnography data, the model goes beyond the traditional stages of sleep and identifies a pattern of disturbed sleep associated with higher mortality.
The sleep stages used in the clinic (N1, N2, N3, paradoxical sleep) cut night into large categories that mask the depth variations within the same stage. A team proposes ECG-SDI, a continuous index of sleep depth calculated from a single electrocardiogram channel, much simpler to collect than a complete polysomnography.
Validated on two cohorts
The two-stage deep learning model was trained and validated on 12,392 night records. Apart from paradoxical sleep, its correlation with the ordered stages reached 0.828 on the internal test set and 0.848 on the external cohort Cleveland Family Study; paradoxical sleep is recognized with a AUROC of 0.954 and 0.977. In the Sleep Heart Health Study, the extracted biomarkers isolate a subtype of "disturbed sleep" associated with lower sleep quality and higher all-cause mortality (adjusted relative risk of 1.20).
The index also appears to be treatment sensitive: in 507 patients registered before and after continuous positive pressure (CPAP), the proportion of normalized deep sleep increases by 6.1% and the sleep load decreases very slightly by 16%. Proof of concept on a portable device suggests a home follow-up, which remains to be evaluated.
Read the full article on the original source
Go to the source →