Abstract Background Continuous follow-up for patients with major depressive disorder (MDD) is essential for treatment decisions and a better prognosis. There remains limited evidence regarding the critical issue of depression variation trajectory prediction using mobile health (mHealth) measures. Moreover, the temporal dynamics of mHealth measures have not been fully modeled in previous studies, and the poor patient adherence to mHealth records poses great challenges to the dynamic feature modeling. Objective This study aimed to examine the contribution of mHealth measures in predicting depression variation trajectory for patients with MDD, with full consideration of the temporal dynamics of mHealth measures. Methods A total of 229 patients with MDD from a multiple-center, prospective cohort were included. A 12-week follow-up was conducted involving the collection of the Hamilton Depression Rating Scale (HAMD-17), along with patient-reported outcomes (Immediate Mood Scaler and Altman Self-Rating Mania Scale) via mobile devices and sleep duration through wearable wristbands. We used functional data analysis to extract dynamic features from the sparse mHealth records, rather than aggregating the data to a single scalar summary measure through collapsing over time. Subsequently, 3 machine learning models were applied to predict the depression variation trajectory classes based on the baseline characteristics and these extracted dynamic features. Results Based on the variation of HAMD-17 scores within 12 weeks, the participants were labeled into 4 classes through the k -means algorithm. The classes included stable decline (n=93), fluctuate decline (n=44), fast decline (n=60), and delayed and fluctuate (n=32), in light of the shape of depression trajectories. With both baseline features and dynamic features of the mHealth measures, accuracy rates for the overall data were 54.35%, 60.87%, and 56.52%, for the stable decline patients were 78.95%, 84.21%, and 73.68%, for the nonstable decline patients were 59.26%, 62.96%, and 70.37% based on the 3 machine learning models, respectively. The results were significantly superior to the prediction obtained without mHealth measures (with an overall accuracy below 50%) and only showed a marginal reduction in accuracy relative to the ideal prediction with assessment obtained from clinical visits. Moreover, in the construction of the most accurate prediction model, dynamic features of the Immediate Mood Scaler, the Altman Self-Rating Mania Scale, and sleep duration emerged as the most influential predictors, ranking first, third, and fourth, respectively, in terms of their relative importance. Conclusions Longitudinal mHealth measures show potential in depression variation trajectory monitoring for patients with MDD even under poor patient adherence. Our work provides practical help in alleviating the follow-up burden for patients with MDD and validates the effectiveness of mHealth measures in clinical applications.
Zhong et al. (2026) studied this question.