This simultaneously records several physiological signals such as brain, heart, muscle and respiratory activity during sleep. So many parameters on which the SleepFM AI was trained based on 585,000 hours of nightly recordings of 65,000 people.
How does SleepFM dissect your sleep?
Table of Contents
Sleep is a complex process based on fine interactions between several physiological systems. Polysomnography captures these interactions through a multimodal analysis of different signals, including brain electrical activity (with the electroencephalogram and electrooculogram), that of the heart (electrocardiography), muscles (electromyography) as well as respiratory signals.
Thus, SleepFM produces representations of sleep that capture its physiological and temporal structure and enable accurate prediction of future disease risk. Researchers estimate that SleepFM could thus identify dozens of diseases well before they manifest, such as Alzheimer’s disease, Parkinson’s disease, certain cancers, cardiovascular diseases or diabetes, based on these 4 types of signals: brain activity, heartbeat, muscle contractions and breathing.
“By leveraging disease codes from electronic medical records (EMR), we developed a framework allowing (AI) to systematically explore the predictive associations between multimodal sleep (and the signals described above, editor’s note) and various conditions,” describe the researchers.
Impressive results from the first night
The performance is remarkable: from a single night of sleep, the model predicts mortality (all causes) with a concordance index of 0.84. The concordance index is an index between 0 and 1. The closer we get to 1, the more accurate the predictions are.
For dementia the index is 0.85, for heart attack 0.81 and for heart failure it is 0.80. The same goes for chronic kidney failure (0.79), stroke (0.78), and atrial fibrillation (0.78). In other words, according to the authors, “the model presents excellent performance for the prediction of death, dementia, heart failure and chronic kidney failure”.
Comparison with other methods
To evaluate its effectiveness, the researchers compared SleepFM to two traditional methods. They explain: “The SleepFM model estimates the risk of all-cause death with greater accuracy than a demographic-based model and an end-to-end PSG model.
This result shows that the pre-training phase makes it possible to identify weak signals linked to the risk of death, particularly in polysomnographic recordings. »
External validation confirms these results. On a cohort completely independent of that used for machine training, SleepFM maintains solid performance.
Characteristic sleep abnormalities
According to the authors, their work establishes a close link between the risk of developing certain pathologies and several parameters linked to sleep: for example, a high wakefulness load (fragmented sleep, marked by numerous micro-awakenings or prolonged periods of wakefulness, etc.), or even paradoxical sleep abnormalities (several years before the onset of Parkinson’s disease), or even nocturnal respiratory disorders and marked hypoxemia: nocturnal respiratory micro-arrests often precede cardiac events. They also appear to be predictive of senile dementia.
SleepFM seems to combine all of these dimensions by integrating respiratory events, cycle fragmentation, arousal load, overall efficiency as well as classic markers associated with cardiovascular, metabolic and systemic pathologies.
It will remain for researchers to generalize these results to the entire population, the study having mainly focused on patients already consulting for sleep disorders.
