Sleep disturbance recorded via wearable sensors predicts depression severity 9 years later.

BACKGROUND

Major depressive disorder (MDD) is prevalent and poses major public health implications. Autonomic nervous system (ANS) dysregulation and sleep disturbances are theorized to be distal risk factors. However, previous research has depended on cross-sectional designs, small predictor sets, and suboptimal methods, limiting temporal inference and predictive accuracy. We thus capitalized on machine learning to identify physiology and sleep predictors of nine-year MDD symptoms.

METHOD: Community adults (N = 1054) participated in a study that included baseline physiological electrocardiogram (ECG) and sleep actigraphy wearable assessments. Clinical interviews were administered to assess for psychiatric symptoms at baseline and nine-year follow-up. Eight ML models were trained to predict MDD severity using 80 baseline variables via a 70-30 train-test split with 5-fold cross-validation with 81 baseline variables to predict MDD severity.

RESULTS

The best model (gradient boosting machine) had 10 variables with strong predictive accuracy in the test set (R2 = 19.8%). Baseline MDD, generalized anxiety, and panic disorder symptoms strongly predicted nine-year MDD severity. Longer total sleep time, lower sleep efficiency, and higher average wake time during sleep phases were key correlates of higher nine-year MDD severity. Other correlates included fewer average sleep bouts and shorter wake times during active phases, as well as nonlinear patterns of wake time length and percentage during rest phases. Physiology ECG variables had limited incremental predictive value.

CONCLUSIONS

Wearable actigraphy-indexed sleep disturbances predicted long-term MDD symptoms beyond baseline severity and ANS dysregulation indices. Combining passive sleep sensors into routine assessments might optimize MDD prevention and treatment.