MACHINE LEARNING DERIVED-INTERPRETATIVE ALGORITHM BETTER DIFFERENTIATES SLEEP AND WAKE EPOCHS AND ESTIMATES SLEEP PARAMETERS FROM WRIST ACTIGRAPHY DATA

Abstract Introduction Several different interpretive algorithms (IAs) are available for scoring actigraphy-obtained body movement data for sleep and wake epochs. Although most have high sensitivity in detecting sleep epochs, they identify wake epochs poorly. We derived a machine learning (ML) based IA that improves differentiation of sleep and wake epoch to better estimate sleep parameters. Methods Forty-one adults (18 females) 26.6±12.0 years old underwent at-home single-night sleep assessment. Motionlogger® Micro Watch Actigraph recorded in zero crossing mode body movement per 30s epoch, with automated sleep scoring by single-channel electroencephalography (EEG) device (Zmachine® Insight+) as reference. The popular Cole-Kripke IA was applied to score body movement time series data of the following combination of current 1, preceding 4, and following 2 minute long epochs. Data of 21 subjects were utilized to train/derive the ML IA (logistic regression), and data of the other 20 subjects were used to test performance of it and the Cole-Kripke IA. Results In reference to the EEG, the Cole-Kripke actigraphy IA showed sensitivity of 0.98±0.02, specificity of 0.48±0.19, and kappa agreement of 0.53±0.16 in detecting sleep epochs, while the ML-derived IA showed corresponding values of 0.90±0.06, 0.71±0.14, and 0.57±0.11. The Cole-Kripke IA, relative to EEG, method significantly (P<0.05) underestimated sleep onset latency (SOL) by 18.0 min and wake after sleep onset (WASO) by 35.1 min, and overestimated total sleep time (TST) by 53.1 min and sleep efficiency (SE) by 9.6%. The ML-derived IA, relative to EEG significantly underestimated SOL by 15.1 min, but comparably (P>0.05) estimated WASO, TST, and SE. Conclusion The ML-derived IA, in comparison to Cole-Kripke IA, when applied to sleep-time wrist actigraphy data significantly better differentiates wake from sleep epochs and better estimates sleep parameters. Support This work was supported by the Robert and Prudie Leibrock Professorship in Engineering at the University of Texas at Austin.