0760 Applying machine learning to examine settling down period for children with and without sensory hypersensitivities

Abstract Introduction Children with sensory hypersensitivities have poorer subjective sleep health than their peers. However, traditional actigraphy variables (e.g., sleep efficiency, sleep duration) do not adequately capture these sleep deficits. In qualitative interviews, caregivers of children with sensory hypersensitivities identified the settling down period prior to sleep as a major family stressor, potentially indicating a novel target for intervention. We applied machine learning techniques to thoroughly characterize and discriminate differences in the settling down period for children with and without sensory hypersensitivities. Methods Children (ages 6-10) with sensory hypersensitivities (n=20) and children without sensory hypersensitivities (n=29) wore the GT9X Actigraph continuously for 2 weeks and caregivers completed daily sleep diaries. Settling down period (caregiver reported start of settling down until actigraphy indicated sleep onset) was isolated for each night and 7 features were extracted from the activity data (mean magnitude, maximum magnitude, kurtosis, skewness, Shannon entropy, standard deviation, interquantile range). Ten-fold cross-validation with random forests were used to determine the accuracy, sensitivity, and specificity of differentiating groups. Results We achieved an 83% accuracy in classifying children with sensory hypersensitivities versus those without hypersensitivities (sensitivity = 84%; specificity = 82%). Feature importance maps showed that the most important feature for differentiating groups was maximum activity count magnitude during the settling down period; children with sensory hypersensitivity had higher maximum bouts of activity during settling down. The hypersensitive group also showed a higher variance in activity during settling down, as demonstrated by greater interquartile range (variance within the time window), standard deviation of activity, and Shannon entropy (the amount of uncertainty in the time window). Conclusion Our novel machine learning analysis successfully uncovered objective features within the settling down period that differentiate children with sensory hypersensitivity from their peers. Our data highlights exciting new potential targets for intervention: children with sensory hypersensitivities have larger and sporadic bouts of activity during their settling down period that clearly set them apart from their peers without hypersensitivities. Support (if any) T32 HL082610, University of Pittsburgh School of Rehabilitation Science Doctoral Award (PI Hartman), Sensory Integration Education PhD Student Grant (PI Hartman).