JMIR Publications

ABSTRACT

Background:

Remote measurement technologies (RMTs) offer a potential solution for monitoring depression symptoms in situ while reducing reliance on self-report assessments. Sleep and physical activity measures derived from wearable devices may capture behavioural patterns relevant to depressions severity, but their incremental value for predicting current depression severity remains unclear.

Objective:

To evaluate the predictive value of wearable-derived sleep and physical activity measures for current depression severity, and to determine whether RMT-derived behavioural subtypes improve prediction compared with individual RMT features and baseline sociodemographic and clinical characteristics.

Methods:

Longitudinal data from the Remote Assessment of Disease and Relapse in Major Depressive Disorder (RADAR-MDD) cohort were used to predict three-monthly depression severity, at three to 36-months of follow-up, measured using the Inventory of Depressive Symptomatology- Self Report. Elastic net and random forest models were compared across different combinations of baseline individual characteristics, individual sleep and physical activity RMT features, and RMT-derived behavioural subtypes. RMT features were derived from data collected during the seven days preceding each IDS-SR assessment. Behavioural subtypes were previously derived from these longitudinal RMT data using a Hidden Markov model, which estimated participants’ subtype membership probabilities at each timepoint. Prediction models were evaluated using nested group five-fold cross-validation and assessed using root mean square error, R2 and mean absolute error.

Results:

Models including baseline depression severity achieved R2 values of approximately 0.6, with RMSE of approximately 10 and MAE of 8. Adding individual RMT features of RMT-derived subtypes provided little improvement in predictive performance. Excluding baseline depression severity reduced R2 to approximately 0.1, while RMSE increased to approximately 15 and MAE to 12, demonstrating the substantial contribution of baseline depression severity to predictive performance.

Conclusions:

Wearable-derived sleep and physical activity measures and their behavioural subtypes provided limited additional predictive value for current depression severity beyond baseline individual characteristics, particularly baseline depression severity. Although these measures did not substantially improve prediction in the present models, they may still capture complementary behavioural information relevant to understanding depression.