Performance of a multisensor ring to evaluate sleep: in-lab evaluation relative to psg and actigraphy: importance of generalized versus personalized scoring

Abstract Introduction Multisensor sleep wearable devices have demonstrated utility for research and relative accuracy for discerning sleep-wake patterns at home and in the laboratory. Additional sensors and more complex scoring algorithms may improve the ability of wearables to assess sleep health. Methods Thirty-six healthy adults completed assessment while wearing the experimental device (Happy Ring), as well as Philips Actiwatch, Fitbit, Oura, and Whoop devices. Evaluations occurred in the laboratory (Alice 6 polysomnogram). The Happy Ring includes sensors for accelerometry, photoplethysmography, electrodermal activity, and skin temperature. Epoch-by-epoch analyses compared the Happy Ring to lab polysomnography, as well as other sleep-tracking devices. Scoring was accomplished using two machine-learning-derived algorithms: a “generalized” algorithm which was static and applied to all users (like those used for other devices) and a “personalized” algorithm where parameters are personalized, dynamic, and change based on data collected across different parts of the night of sleep. Results Compared to in-lab polysomnography, the generalized algorithm using data from the Happy Ring demonstrated good sensitivity (94%) and specificity (70%). The personalized algorithm also performed well with good sensitivity (93%) and specificity (83%). Other devices also demonstrated good sensitivity, ranging from 89% (Fitbit) to 94% (Actiwatch); specificity however, was more variable, ranging from 19% (Actiwatch) to 54% (Whoop). Overall accuracy was 91% for generalized and 92% for personalized, compared to 88% for Oura, 86% for Whoop, 84% for Fitbit, and 85% for Actiwatch. Measurement of sleep stage accuracy was 67%, 85%, and 85% for light, deep, and REM sleep, respectively, for the Happy generalized algorithm. For the Happy personalized algorithm, accuracy for sleep stages were 81%, 95%, and 92%, for light, deep and REM sleep, respectively. Post-hoc analyses showed that the Happy personalized algorithm demonstrated better specificity than all other modalities (p<0.001). Kappa scores were 0.45 for generalized and 0.68 for personalized, compared to 0.32 for the Oura Ring, 0.32 for Whoop Strap, and 0.37 for Fitbit wristband. Conclusion The multisensory Happy ring demonstrated good sensitivity and specificity for the detection of sleep in the laboratory. The personalized approach outperformed all others, representing a potential innovation for improving detection accuracy. Support (If Any) Dr. Grandner is supported by R01DA051321 and R01MD011600. This work was supported by Happy Health, Inc.