Estudo da Stanford Revela IA para Prever Riscos de Doença a Partir de Dados de Sono

Stanford Researchers Build AI to Predict Disease Risk from Sleep Data

Recent studies conducted by researchers at Stanford University have introduced an artificial intelligence model called SleepFM, capable of predicting disease risks in individuals from a single night of sleep data. This remarkable advancement promises to revolutionize how we approach preventive health care by providing insights into over 130 medical conditions, including heart disease, various cancers, and mental disorders.

What is SleepFM?

The SleepFM model utilizes data obtained through polysomnography (PSG), which captures multiple biological signals, including:

Brain waves (EEG)
Heart activity (ECG)
Respiration and respiratory effort
Muscle activity (EMG)
Eye movements
Blood oxygenation levels

Instead of analyzing these metrics in isolation, the model treats them as an integrated physiological set, allowing for the identification of patterns that may remain unnoticed in traditional analyses.

Model Integration and Training

The model was trained using approximately 585,000 to 600,000 hours of sleep data collected from around 65,000 individuals. This data was segmented into 5-second intervals, similar to the methodology used in language models, which enables the AI to learn the normal interactions between different bodily systems during sleep.

Prediction Mechanism

SleepFM identifies subtle discrepancies or “out-of-phase” patterns that may indicate underlying pathological processes. The AI employs a technique known as contrastive learning, where a certain signal is removed, and the model attempts to reconstruct it using the remaining signals. This methodology enhances its ability to detect anomalies.

Performance in Disease Prediction

Out of more than 1,000 disease types examined, 130 conditions demonstrated a reasonable level of predictive accuracy. Notable performance metrics include:

Disease	Concordance Index (C-index)
Parkinson’s Disease	0.89
Dementia	0.85
Hypertensive Heart Disease	0.84
Heart Attack	0.81
Prostate Cancer	0.89
Breast Cancer	0.87
Overall Mortality	0.84

A concordance index above 0.8 indicates that the AI can accurately classify individuals based on future risk in over 80% of cases.

Time Horizon and Limitations

The model’s effectiveness is augmented by access to electronic health records spanning up to 25 years of follow-up. Thus, the predictions can reveal risks long before actual clinical diagnoses, though the exact timelines might vary depending on the specific condition.

However, it’s crucial to recognize that this research is still in its early stages. The model was developed based on patients in sleep clinics, suggesting further validation will be needed before it can be applied to the general public. It’s also important to note that while the model provides risk stratification, it does not guarantee a definitive diagnosis.

Future Directions

The Stanford team plans to refine SleepFM and expand its application to data from consumer devices, such as wearables. This innovation could potentially make sleep-based preventive health screening more accessible, provided the model’s performance remains robust outside specialized laboratory settings.

Conclusion

The development of SleepFM by Stanford researchers marks a significant step towards leveraging sleep data for disease prediction. If validated and successfully implemented, this system could not only foresee health risks but also transform our understanding of wellness and medical prevention. This research highlights not only the complexity of the data captured during sleep but also the potential of AI technologies to reshape preventive medicine in the future.

Sources

Share this post

Like this post? Share it with your friends!