automatic sleep detection

[nealgravindra]

Action items

• [ ] reproduce and implement Palotti et al. 2019 scripts for sleep detection (even simple heuristics)
• [ ] visualize sleep detection results and remark upon baselines for dataset by using survey data to inform performance
• [ ] play around with publicly available

Models

For count-based sleep detection:

• Cole-Kripke (7min time window)
• Sadeh (11min time window)
• Oakley (5min time window)
• Borazio (low SD/s lasting for 10min): heuristic

For acceleration vectors:

• van Hees (sleep based on time segments where est. angle of accelerometer rel. to gravity does not change beyod 5deg. for 5min): popular because it facilitated easier interpretation to approaches based on magnitude of acceleration
• Trevenen et al. (ML to extract features from acceleration magnitude and use these as input to HMM to classify sleep v. wake; however had poor classification performance on sleep stage)
• Sundarajan et al. 2021: (use RF on feature extracted from GENEActiv accelerometers, and labels come from technician scoring PSG data; also have hierarchy of classification: wear v. non-wear --> if wear, sleep v. wake --> if sleep, sleep stage classification)

Datasets

This is key, since many tasks seem to rely on either heuristics or supervised approaches with the label derived from sleep diaries or polysomnography, which is expensive and limited in scale

• MESA Sleep study (Mutli-Ethnic study of Atherosclerosis with synchronized polysomnography data and actigraphy; benchmarks with dataset introduced in Palotti et al. 2019)
• sleepdata.org and 3 datasets (total ~20k people) with polysomnagraphyu and actigraphy data (see specific page here)

Conclusions

• Palotti et al. 2019 is a very well-done study and the most comprehensive and scientifically sound. It far exceeds Sundarajan et al. 2021 in quality. Palotti et al. also establish a number of baselines to beat, in particular, InceptionTime and transformer-based InceptionTime (+Time2Vec) could improve performance.
• Outstanding problems include transfer-ability of models to different datasets and different devices

References

• Sundarajan et al. 2021
• Palotti et al. 2019

[nealgravindra]

For now, doing something simple with pyactigraphy