Predicting Depression From Smartphone Behavioral Markers Using Machine Learning Methods, Hyperparameter Optimization, and Feature Importance Analysis: Exploratory Study

Background Depression is a prevalent mental health challenge. Current depression assessment methods using self-reported and clinician-administered questionnaires have limitations. Instrumenting smartphones to passively continuously collect moment-by-moment data sets quantify human behaviors has the potential augment current for early diagnosis, scalable, longitudinal monitoring of depression. Objective The objective this study was investigate feasibility predicting with quantified from smartphone sets, identify that can influence Methods Smartphone 8-item Patient Health Questionnaire (PHQ-8) assessments were collected 629 participants in an exploratory over average 22.1 days (SD 17.90; range 8-86). We 22 regularity, entropy, SD behavioral markers data. explored relationship between features correlation bivariate linear mixed models (LMMs). leveraged 5 supervised machine learning (ML) algorithms hyperparameter optimization, nested cross-validation, imbalanced handling predict Finally, permutation importance method, we identified influential Results Of at least 56 countries, 69 (10.97%) females, 546 (86.8%) males, 14 (2.2%) nonbinary. Participants’ age distribution as follows: 73/629 (11.6%) aged 18 24, 204/629 (32.4%) 25 34, 156/629 (24.8%) 35 44, 166/629 (26.4%) 45 64, 30/629 (4.8%) 65 years over. 1374 PHQ-8 assessments, 1143 (83.19%) responses nondepressed scores (PHQ-8 score <10), while 231 (16.81%) depressed ≥10), based on cut-off. A significant positive Pearson found screen status–normalized entropy (r=0.14, P<.001). LMM demonstrates intraclass 0.7584 association (β=.48, P=.03). best ML achieved following metrics: precision, 85.55%-92.51%; recall, 92.19%-95.56%; F1, 88.73%-94.00%; area under curve receiver operating characteristic, 94.69%-99.06%; Cohen κ, 86.61%-92.90%; accuracy, 96.44%-98.14%. Including group gender predictors improved performances. Screen internet connectivity most Conclusions Our findings demonstrate indicative be unobtrusively sensors’ Traditional augmented diagnosis monitoring.