A Multi-Stage Framework for Refining Infant Daytime Sleep-Wake Labels from Wearable Accelerometer Data.

Sleep is essential for infants' physical and cognitive development. Unlike older children or adolescents, infants sleep longer durations, including multiple daytime naps. While nighttime sleep is easier to detect due to its extended periods, identifying daytime sleep is more challenging due to its short, fragmented nature and its similarity to idle wakefulness. Moreover, parent-reported sleep estimations are prone to error as continuous monitoring is often impractical. To address these limitations, we developed a Multi-Stage Sleep-Wake (MSW) classification approach using triaxial accelerometer data collected from wearable devices placed on infants' ankles and waists over multiple days. This method systematically refines and classifies sleep and wake states through a series of analytical steps. This systematic process generates refined proxy sleep-wake labels that account for behavioral overlaps and correct mislabeled idle periods. We trained and validated multiple machine learning models using these labels and compared the results to parent annotated labels. Models trained using MSW-derived labels achieved 96.6% accuracy using Random Forest classifier compared to 72% using parent-reported labels. These findings demonstrate that the MSW framework produces a more consistent set of proxy sleep-wake annotations for model development, although the derived labels were not validated against an independent reference standard. Furthermore, the proposed MSW framework may serve as a practical label-refinement methodology for improving noisy caregiver-reported sleep annotations in future wearable-based infant sleep studies where independent sleep labels are unavailable.