Machine learning approaches in sleep and circadian research

Machine learning (ML) offers some of the most cost-effective methods for extracting useful insights from large data sets. Availability of large data sets and tools for modular, scalable, reproducible, open, shareable data analytic workflows, allow researchers to rapidly train, rigorously validate and share predictive models in health research. These developments offer unprecedented opportunities for overcoming some of the limitations in the analyses of sleep- and circadian-related data, including data collected using the field's “gold standard” clinical monitoring techniques, such as polysomnography (PSG) and evaluation of ML-trained models, e.g., for estimating sleep parameters. Sleep and circadian rhythms researchers increasingly use multi-modal monitoring techniques, often with humans evaluating the data who make inferences about the collective integration of those signals when “scoring” or annotating the data (e.g., with sleep vs. wake states, sleep stage, etc. We support and extend the rationale supporting the AASM's position statement that the multi-modal and complex nature of data collected in clinical sleep monitoring is “uniquely positioned to benefit from the use of artificial intelligence”. We emphasize some specific applications to sleep and circadian rhythms research, and argue for the expanding potential of ML and Artificial Intelligence (AI) in the broader scientific approach to sleep and circadian research. We also underscore some of the factors that must be addressed to ensure methodologically rigorous applications of ML in this area.