What goes on when the lights go off? Using machine learning techniques to characterize a child's settling down period

Abstract

Objectives: Current approaches to objective measurement of sleep disturbances in children overlook the period prior to sleep, or the settling down time. Using machine learning techniques, we identified key features that characterize differences in activity during the settling down period that differentiate children with sensory sensitivities to tactile input (SS) and children without sensitivities (NSS).

Methods: Actigraphy data were collected from children with SS (n = 17) and children with NSS (n = 18) over 2 weeks (a total of 430 evenings). The settling down period, indicated using caregiver report and actigraphy indices, was isolated each evening and seven features (mean magnitude, maximum magnitude, kurtosis, skewness, Shannon entropy, standard deviation, and interquartile range) were extracted. 10-fold cross-validation with random forests were used to determine accuracy, sensitivity, and specificity of differentiating groups.

Results: We could accurately differentiate groups (accuracy = 83%, specificity = 83%, sensitivity = 84%). Feature importance maps identify that children with SS have higher maximum bouts of activity (U = -2.23, p = 0.026) during the settling down time and a higher variance in activity for the children with SS (e.g., interquartile range, Shannon entropy) that sets them apart from their peers.

Conclusion: We present a novel use of machine learning techniques that successfully uncovered differentiating features within the settling down period for our groups. These differences have been difficult to capture using standard sleep and rest-activity metrics. Our data suggests that activity during the settling down period may be a unique target for future research for children with SS.

Keywords: actigraphy; children; machine learning; sensory processing; sensory sensitivity; settling down; sleep; sleep onset delay.

FIGURE 1

Average AUC change for features.