Pediatric-trained LSTM improves actigraphy sleep-wake classification in children

Researchers evaluated whether machine-learning models trained on pediatric data could classify sleep and wake from raw wrist-accelerometer recordings. Sixty-five children aged 5.3–17.7 years underwent overnight polysomnography while wearing a GENEActiv device. The team divided acceleration data into 30-second epochs, aligned them with physician-scored sleep-wake data, and used leave-one-subject-out cross-validation to assess eight models. Comparators included GGIR algorithms and an adult-trained random forest classifier.

Across 560.1 hours of recordings, 74.4% of epochs were scored as sleep. The pediatric-trained local-global long-short-term memory (LSTM) classifier performed best, with balanced accuracy of 0.85, sensitivity of 0.88, specificity of 0.83, ROC-AUC of 0.95, and Cohen’s kappa of 0.67. Its performance exceeded that of the adult-trained classifier and the GGIR algorithms. The LSTM underestimated overall sleep duration by an average of 25 minutes, with no proportional bias reported.

The findings support further evaluation of pediatric-trained models for actigraphy-based sleep-wake classification. The study assessed algorithm performance against in-lab polysomnography in a sample of 65 children; the abstract does not establish how well the model generalizes to other devices or real-world settings.