Poor sleep can leave you tired the next morning. But in the near future It may become an early warning sign of disease developing in the body for years to come. By a team of researchers from Stanford Medicine developed a model SleepFM which can analyze physical data from just one night of sleep To predict the risk of developing more than 100 diseases, many years before disease symptoms appear.
SleepFM Trained with high standard sleep monitoring data or Polysomnography (PSG) than 600,000 hours from the patient 65,000 people It records vital signals of the body such as brain waves (EEG), electrocardiogram (ECG), breathing, and body movement.
Emmanuel Mignot Professor of Sleep Medicine from Stanford specify that “Every time we study sleep We collected a huge amount of data from the body. During the 8 hours the subjects were lying still, we were looking at almost every system in the body working at the same time. The information obtained is therefore very detailed and dense.”
Until recently, these data were only partially used in sleep medicine. But with advances in AI, researchers are starting to be able to analyze and understand this huge amount of data. This research is the first time that AI has been used to analyze sleep data on such a large scale.
AI that learns the ‘language of sleep’
design research team SleepFM Let it be Foundation Model In the same concept as ChatGPT, it learns from huge amounts of text data. and apply the knowledge gained to various forms of work But instead of learning from written language This model learns from ‘Language of Sleep’
Information from inspection Polysomnography It is divided into 5 second intervals which function similarly to ‘vocabulary’ where large-scale language models are used to learn from text It allows AI to systematically learn the patterns and relationships of sleep signals from many body systems.
Including using techniques Leave-one-out Contrastive Learning It hides some data in order to train the AI to predict missing signals from other data. This process helps the model understand the in-depth interplay between the body’s organs.
James Zou Associate Professor of Data Science explains that the accuracy of the model does not come from looking at data from a single organ alone. But it comes from taking signals from all systems and analyzing them together. Especially when found ‘Uncoordinated functioning of organs’ For example, the brain is in a state of deep sleep. But the heart still shows a pattern of wakefulness. This is an accurate indicator of long-term health risks.
Prognostic accuracy
When SleepFM was tested against a longer history of medical records, 50 years of Stanford Sleep Medicine Center The results show that SleepFM Data from more than 1,000 disease groups in medical records were analyzed and found that 130 disease groups could be predicted.Accuracy at a practical level, e.g.
- Parkinson’s disease (accuracy 89%)
- Prostate cancer (accuracy 89%)
- Dementia (accuracy 85%)
- Heart failure (accuracy 84%)
- Risk of death (accuracy 84%)
The research team says they are continuing to develop SleepFM to increase its accuracy in future prognoses. It may be combined with additional data from wearables.
Researchers consider that The success of SleepFM may change the role of sleep monitoring. from disease-specific diagnostic tools Going to predictive health screening That helps watch out for serious diseases before they show symptoms. and may become one of the important tools of preventive health care in the future.
Reference: Stanford Medicine