JMIR Publications

ABSTRACT

Background:

Continuous follow-up for patients with major depressive disorder (MDD) is essential for treatment decision and 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 modelled in previous studies, and the poor patient adherence of mHealth records poses great challenges to the dynamic feature modelling.

Objective:

This study aimed to examine the contribution of mHealth measures in predicting depression variation trajectory for MDD patients, with full consideration of the temporal dynamics of mHealth measures.

Methods:

A total of 229 MDD patients from a multiple centers, prospective cohort were included. A 12-week follow-up period involved the collection of Hamilton Depression Rating Scale for Depression (HAMD-17) scale assessments, along with regular patient-reported outcomes (PROs) via mobile devices and sleep data acquisition through wearable wristbands. We utilized a functional data analysis method to extract the dynamic features of the sparse records of mHealth measures. Decision tree, random forest, XGBoost and neural networks were applied in the prediction of the depression variation trajectory class using baseline features, dynamic features of mHealth measures, and HAMD-17 and Hamilton Anxiety Scale (HAMA) scores up to different weeks.

Results:

Based on the variation of HAMD-17 scores within 12 weeks, the participants were labeled into four classes through k-means algorithm. The classes were named Stable decline (n = 93), Fluctuate decline (n = 44), Fast decline (n = 60), Delayed and fluctuate (n = 32) based on the shape of depression trajectories. With the sequential augment of HAMD-17 and HAMA scores in the models, the classification accuracy improved correspondingly. With only baseline features, overall accuracy was below 50% across all models. The incorporation of week 2 clinical assessments yielded significant accuracy improvements across models (neural networks: 39.02% to 59.02%), with subsequent progressive enhancements observed through week 8 (neural networks: accuracy 74.74%). Classification performance for the Stable decline category remained consistently robust (≥70% accuracy from week 4), while non-Stable decline identification demonstrated particularly pronounced gains (neural networks: 79.81% accuracy).

Conclusions:

Longitudinal mHealth measures show potential in depression variation trajectory prediction for MDD patients even under poor patient adherence. Our findings suggest that tracking PROs and digital phenotypes, combined with HAMD-17 and HAMA scores up to week 2, may provide clinically meaningful information for predicting depression variation patterns. Our work provides practical help in alleviating the follow-up burden for MDD patients and validates the effectiveness of mHealth measures in clinical applications.