Machine learning sleep duration classification in Preschoolers using waist-worn ActiGraphs.

OBJECTIVE

To create a sleep duration classification technique for waist-worn ActiGraph accelerometers in preschool-aged children.

METHODS

Children wore ActiGraph wGT3X-BT accelerometers on their right hip for 7 days (24 h/day). Ground truth nap, sleep, and wake were estimated through visual inspection of accelerometer data, guided by sleep log-sheets and previously published visual inspection heuristics. Raw accelerometer data (30Hz) were used to generate 144 features aggregated to 1-min epochs. Machine learning classification (ie, Random Forest and Hidden Markov Modeling [HMM]) predicted nap, sleep, and wake. A simplified prediction formula was also created using features (n = 10) with the highest mean decrease in Gini index during training of Random Forests, and temporally smoothed with rolling median calculations.

RESULTS

Children (n = 89, mean age = 4.5 years, 67% boys) contributed >600,000 min of accelerometer data. Overall classification accuracy of the Random Forest and HMM classifier was 96.2% (95%CI: 96.1, 96.2%), with a Kappa score of 0.93. Additionally, overall classification accuracy for the temporally smoothed simplified formula was 93.7% (95%CI: 93.6, 93.7%) with Kappa = 0.87. Nap prediction accuracy was 99.8% for the final machine learning model, and 86.1% for the simplified formula. For participant-level daily summaries, generally small but statistically significant differences were found between machine learning and ground truth behaviour predictions, whereas non-significant differences were found between the simplified formulas and ground truth predictions.

CONCLUSIONS

Predictions for both machine learning and the simplified formula had almost perfect agreement with visual inspection ground truth measurements. Future research is needed to confirm these findings using polysomnography ground truth sleep measurements.