Abstract
Background: Accurate assessment of sleep is vital, but the gold standard, polysomnography, is costly and impractical for large-scale studies. An alternative is wearable accelerometers, which reduce participant burden and avoid biases associated with subjective reporting e.g. recall biases. This study aimed to develop and validate a method for estimating time in bed (TIB), total sleep time (TST), sleep efficiency (SE), sleep onset latency (SOL), and wake after sleep onset (WASO) utilizing machine learning applied to thigh-accelerometry data.
Methods: Data on 309 nights from 134 children aged 4–17 years was used to develop a method utilizing two machine learning models applied to data from thigh-worn accelerometers to estimate sleep metrics. Inputs were collected simultaneously from the Zmachine Insight + and raw data from thigh-worn accelerometers, trained using k-fold cross-validation and validated in a participant-disjoint internal hold-out set. The method was then externally validated against polysomnography in an independent sample of 136 children aged 8–16 years.
Results: The independent validation showed overestimations of 28.0 min for bedtime and 11.2 min for wake time, with ICC of 0.59 and 0.55. TIB and TST were overestimated by 14.0 and 3.3 min with ICC of 0.59 and 0.56, respectively. The correlation for estimating SE, SOL and WASO was weak with ICC of 0.21, 0.01 and 0.04, respectively.
Conclusions: This method demonstrated sufficient accuracy for assessing bedtime, wake time, TIB and TST at the group level when validated in an independent sample against polysomnography, although wide limits of agreement suggest limited precision for individual-level assessments. Low agreement for SE, SOL and WASO indicated insufficient accuracy of the method for these metrics.