This study used data from an epidemiological, noninterventional, cohort study conducted between February 2019 and April 2020 at 4 psychiatric hospitals or units in general hospitals in China. The study was approved by the Independent Medical Ethics Committee Board of Beijing Anding Hospital and the other 3 sites (ethical approval no. 2018-119-201917FS-2). All patients provided written informed consent to participate in the study. To ensure confidentiality, each participant was assigned an identification number. Moreover, participants received ¥100 (US $15.5) for each follow-up visit.

Study inclusion and exclusion criteria have been reported [20]. Generally, inclusion criteria for the study included outpatients (1) aged 18‐60 years and (2) diagnosed with MDD according to the Diagnostic and Statistical Manual of Mental Disorders (fourth edition) criteria. Exclusion criteria included patients who had axis I primary psychiatric diagnosis other than MDD or had substance abuse. Patients in this cohort were followed up at 2, 4, 8, and 12 weeks.

A total of 229 patients with MDD were included in the study (Figure 2), with 78 males (34.1%), 151 females (65.9%), and a mean age of 28.7 years. According to the variation of HAMD-17 scores, the participants were labeled as stable decline, fluctuate decline, fast decline, and delayed and fluctuate, respectively. The number of classes was selected as k=4 with the consideration of both silhouette coefficients (0.182) and clinical significance. Clinical significance refers to the capacity of the chosen clusters to distinguish 4 clinically recognizable patterns of symptom reduction over 12 weeks, categorizing the longitudinal patient responses into stable decline, fluctuate decline, fast decline, and delayed and fluctuate trajectories.

Figure 3 shows the variation of the mean HAMD-17 scores from baseline to week 12 for each class (refer to Multimedia Appendix 1 for HAMD-17 scores figures when k=2,3,5), and Figure S4 in Multimedia Appendix 1 shows that patients within the same class exhibited the same temporal trend of depression severity. In specific, the stable decline class contained 93 (40.6%) patients with MDD, who were more likely to improve steadily during the study. There were 44 (19.2%) patients with MDD in the fluctuate decline class. It is shown that the HAMD-17 scores of participants in the fluctuate decline class might go through an upward fluctuation before week 2 and begin to decrease afterwards. The fast decline class included 60 (26.2%) patients with MDD, and the HAMD-17 scores of these patients showed a rapid decline before week 2. Moreover, 32 (14%) patients with MDD were assigned to the delayed and fluctuate class. It is evident that the HAMD-17 scores of these patients showed a delayed decline and were prone to exhibit a drastic rise at week 8. The symptomatic rebound at week 8 indicates a critical clinical juncture for evaluating treatment efficacy in depression [42]. This pattern suggests a unique phenotype of therapeutic instability where early progress fails to persist. Such findings emphasize the importance of extended clinical monitoring to identify patients at risk of midstage relapse.

Table 1 reports the demographic and clinical characteristics of the patients with MDD in each class. It can be observed that the HAMD-17 and HAMA scores, as well as the changes in HAMD-17 scores, were significantly different among these 4 classes for each week. In addition, the average IMS scores during the follow-up were also significantly different among these 4 classes. Moreover, there was no significant difference in sex, age, BMI, and baseline Quality of Life Enjoyment and Satisfaction Questionnaire—Short Form, Sheehan Disability Scale, IMS, and ASMS scores among these 4 classes. Refer to Multimedia Appendix 1 for the figures of mean curves of IMS scores, ASMS scores, and sleep duration for each class, and for the demographic characteristics of patients in each participating center.

Figure 4 and Table S2 in Multimedia Appendix 1 show the classification results obtained from various input information by different machine learning approaches. It is shown that by using baseline features and FPC scores of the PROs and digital phenotype records (scenario 2), decision tree, random forest, and XGBoost all give reasonable classification accuracy for the overall data (54.35%, 60.87%, and 56.52%), stable decline patients (78.95%, 84.21%, and 73.68%), and nonstable decline patients (59.26%, 62.96%, and 70.37%). On the other hand, when the models were constructed without dynamic features of the digital mHealth measures (scenario 1), all considered machine learning approaches show poor classification efficacy, which indicates the great importance of digital mHealth measures in the monitoring of depression variation. Further, comparing the predictive performance in scenario 2 with that in scenarios 3 and 4, though the classification accuracy was elevated, the improvement was not as obvious as that achieved by changing from scenario 1 to scenario 2. Moreover, despite a marginal reduction compared to the ideal clinical scenarios, scenario 2 offers practical advantages in terms of scalability and reduces assessment burden for continuous follow-ups.

Figure S5 and Table S3 in Multimedia Appendix 1 show the variable importance for decision tree, random forest, and XGBoost in the prediction of depression variation patterns using baseline features and FPC scores of the PROs and digital phenotype records (scenario 2). For both random forest and XGBoost, which yielded more accurate classification, the dynamic feature of IMS score records was of paramount importance in the model construction. Moreover, FPC scores of the ASMS score and sleep duration also played a dominant role in the modeling, closely following the baseline HAMD-17 score. Table S4 in Multimedia Appendix 1 reports the P values of the permutation test, which confirm the significance of the contribution of both baseline HAMD-17 score and mHealth measures. The variable importance results demonstrated the great contribution of digital mHealth measures in the prediction of depression variation.

Table S5 in Multimedia Appendix 1 reports the classification results of decision tree, random forest, and XGBoost with the use of baseline features and the mean values of IMS score, ASMS score, and sleep duration, rather than the FPC scores of them (scenario 2). Compared with introducing the dynamic features of the mHealth measures, using the mean values led to inferior prediction, especially for decision trees (overall data: 54.35% vs 47.83%; stable decline: 78.95% vs 47.37%; nonstable decline: 59.26% vs 59.26%) and random forests (overall data: 60.87% vs 50.00%; stable decline: 84.21% vs 78.95%; nonstable decline: 62.96% vs 55.56%). The results implied the necessity of considering the temporal dynamics of the digital mHealth measures.

This study evaluated the contribution of digital mHealth measures, including PROs and digital phenotype, to the prediction of the variation of depression within 12 weeks for patients with MDD. Based on the variation of HAMD-17 scores at 5 visits within 12 weeks, each patient was labeled as stable decline, fluctuate decline, fast decline, or delayed and fluctuate. The study used functional data analysis methods to address the severe sparsity of PROs and digital phenotype records caused by poor patient adherence, thereby effectively extracting their dynamic features. The classification results showed that by combining the baseline features with the dynamic information of PROs and digital phenotypes, the depression variation pattern can be identified with reasonable accuracy. Overall, our study demonstrated that the dynamic features of mHealth measures extracted by the functional data analysis method can be an effective alternative to clinical visits for depression variation monitoring. The findings can alleviate the burdens for both patients and doctors.

For the detection of latent trajectory classes regarding depression severity, growth mixture modeling was widely used in the previous studies [43-45]. However, growth mixture modeling was more likely to capture the similarity in the magnitude of depression severity, rather than its changing pattern. In our study, the patients with MDD were classified into 4 classes based on k-means clustering through the HAMD-17 score differences. Within each class, the patients exhibited the same temporal trend of depression severity. The classes were further named based on the dynamic characteristics of the depression symptom trajectories. In specific, stable decline implies a stable improvement of depression symptoms, fluctuate decline indicates a fluctuation of depression symptoms in the early stage and an alleviation at the end, fast decline represents a high baseline severity and a rapid improvement of depression symptoms, while delayed and fluctuate means a delayed improvement and sharp variation of depression symptoms.

Our findings revealed that the dynamic features of digital mHealth measures can reflect the depression variation of patients with MDD. By recording the trajectories of IMS scores, ASMS scores, and sleep duration using the smartphone and wristband, the dynamic characteristics of depression severity can be evaluated without strong dependence on continuous tracking of HAMD-17 scores. Moreover, results from the variable importance analysis highlighted the contribution of the mHealth measures in the depression variation prediction.

Further, the sensitivity analysis indicated that taking into account the temporal dynamics of the mHealth measures is beneficial. In our study, the dynamic features of PROs and digital phenotypes were fully characterized through a functional data analysis method, rather than aggregating the records to a single scalar summary measure (eg, mean or median) by collapsing over time. Functional data analysis is popular in the study of continuously recorded data, such as physical activity data [46] and electroencephalography data [35], which supports the use of functional data analysis methods in our research. However, classical functional data analysis methods are only suitable for densely collected data. To extract dynamic features of the sparsely collected PROs and digital phenotypes, the PACE method was applied, which is a commonly used method for the principal component analysis of sparsely and irregularly observed functional data [36]. Collectively, the benefits of using functional data analysis methods can be summarized in two aspects: (1) avoiding the loss of information related to temporal dynamics of the PROs and digital phenotype records and (2) conquering the problem of sparse observations caused by poor patient adherence without excluding any participants with a low observation size (or in other words, with a high percentage of missing observations).

Previous studies have demonstrated the potential of mHealth measures in predicting symptom severity of patients with MDD [6,23,47]. To further monitor the depression stability, Bai et al [20] defined two types of variation patterns, steady and swing, based on the patient-reported Patient Health Questionnaire-9 scores. However, such PRO records cannot substitute for the gold-standard HAMD-17 scores evaluated by clinicians in the longitudinal monitoring of patients with MDD. In our study, the more credible clinician-reported scores were used for outcome labeling, and PROs were incorporated as predictive factors, underscoring the need for more sophisticated use of PRO records. Moreover, Ikäheimonen et al [48] also considered variation of depression by partitioning the participants into 4 classes: declines, increases, remains depressed, and remains nondepressed. Although their model based on smartphone behavioral data achieved an overall accuracy of 75%, the predicting efficacy for classes with state changes (declines and increases) was invalid with a precision lower than 50%. Compared with their work, we refined the depression variation labels to facilitate the detection of specific temporal patterns, and our model yielded a more robust predictive performance for patients with fluctuating clinical courses. This improvement likely stems from our application of the functional data analysis to the mHealth records, which can effectively characterize the continuous underlying transition patterns that traditional summary statistics might overlook.

The core contribution of our study lies in the methodological integration of dual temporal dynamics for depression severity and mHealth measures, and the capacity to handle extremely sparse mHealth records. First, our definition of the outcome labels effectively characterized the temporal evolution of depression severity, enabling the identification of the inherent heterogeneity in depression progression over time. Moreover, the dynamic features of mHealth measures were extracted to predict the changing patterns of depression trajectory, forming an innovative dual-dynamic framework that is more practical in real-world psychiatric monitoring. Last but not least, our study made it feasible to obtain an acceptable depression monitoring using mHealth records with great irregularity and sparsity, which significantly enhances the ecological validity of our model and mitigates the adherence-related barriers that hinder clinical application of mHealth measures.

In recent years, high-resource biological markers derived from neuroimaging [49] and electrophysiology [50] have gained prominence in predicting antidepressant response due to their insights into underlying mechanisms. Compared with these markers, low-cost mHealth measures possess the ability of high-frequency, continuous recording within a naturalistic environment, facilitating scalable real-world monitoring and the dynamic capture of longitudinal symptom variations that static biomarkers may overlook. Given the complementary strengths, the integration of high-resource biological markers with real-world mHealth data represents a promising avenue for personalized psychiatry that warrants further investigation.

The strengths of our study can be summarized as follows. First, this study contained a long period of continuous monitoring for a large number of patients with MDD. Second, our study overcame the missing and sparse problems that universally existed in natural observations with the use of functional data analysis methods, which also facilitated the characterization of the temporal dynamics of PROs and digital phenotype records. Third, our study effectively reduced the dependence of patient adherence and eased the heavy burden of clinical follow-up in the prediction of depression variation trajectory.

Our study still has several limitations. One limitation is that only the IMS score, ASMS score, and sleep duration were continuously monitored and were used in the prediction of depression severity. Additional collection of more indicators of PROs and digital phenotype would be beneficial [27]. The other limitation is that only HAMD-17 scores were used in the construction of the labels of depression variation, which may be insufficient. Therefore, collecting more indicators of depressive symptoms and putting them together would be advantageous for the characterization of the change of depression severity. In addition, as the entire sample was used in the k-means clustering analysis to ensure a robust definition of clinical phenotypic labels, an implicit data leakage was introduced into the model assessment. So, the results, which may be inclined toward optimism, should be interpreted with caution. Moreover, constrained by the limited and imbalanced sample sizes of the participating centers, the predictive performance exhibited variability across heterogeneous clinical environments. This suboptimal performance underscores the inherent challenges of model generalizability. Thus, further research is required to enhance the robustness of the model when applied to diverse clinical settings. Last, while the predicted depression variation pattern can guide further treatment for the patient with MDD, it was not able to provide an earlier warning within 12 weeks, which is worth further research.

Our study demonstrated the potential of mHealth measures, as an alternative to clinical visits, in the monitoring of depression variation, even under poor patient adherence. The results suggested that dynamic features of PROs and digital phenotypes, combined with baseline clinical features, can provide clinically meaningful information for depression variation patterns. Our findings help reduce dependence on strict patient adherence and alleviate the burdens of frequent clinical follow-ups, which facilitates a more practical monitoring of depression trajectory variations.