JMIR Publications

ABSTRACT

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 and continuously collect moment by moment datasets to quantify human behaviours that have the potential to augment current depression assessment methods for early diagnosis, scalable, and longitudinal monitoring of depression.

Objective:

The objective of this study is to investigate the feasibility of predicting depression with human behaviours quantified from a smartphone datasets, and to identify behaviours that can influence depression.

Methods:

Smartphone datasets and self-reported eight-item Patient Health Questionnaire (PHQ-8) depression assessments were collected from 629 participants in an exploratory longitudinal study over an average 22.1 days (SD =17.90, min= 8, max=86). We quantified 22 regularity, entropy, and standard deviation behavioural markers from the smartphone usage data. We explore the linear relationship between the behavioural features and depression using correlation and bivariate linear mixed models (LMM). We leverage 5 supervised machine learning (ML) algorithms with hyperparameter optimization, nested cross-validation, and imbalanced data handling to predict depression. Finally, with the Permutation Importance method, we find influential behavioural markers in predicting depression.

Results:

Of the 629 participants from at least 56 countries, 10.96% were females, 86.80% males, 2.22% non-binary. For participants’ age distribution; 11.61% were between 18–24 years, 32.43% 25–34, 24.80% 35–44, 26.39% 45–64 and 4.77% were 65 years and over. Of the 1374 PHQ-8 assessments 83.19% were non-depressed, 16.81% were depressed, based on PHQ-8 cut off. Significant positive Pearson’s correlation was found between screen status normalised entropy and depression (r=0.14, P<.001). LMM demonstrates intra-class correlation of 0.7584 and significant positive association between screen status normalised entropy and depression (beta=.48, P=0.03). The best ML algorithms obtained precision (85.55%–92.50%), recall (92.19%–94.38%), F1 (88.73%–93.41%), area under the curve receiver operating characteristic AUC (94.68%–98.83%), Cohen’s kappa (86.61%–92.21%), and accuracy (96.44%–97.97%). Including age group and gender as predictors improved the ML performances. Screen and Internet connectivity features were the most influential in predicting depression.

Conclusions:

Our findings demonstrate that behavioural markers indicative of depression can be unobtrusively identified from smartphone sensors’ data. Traditional assessment of depression can be augmented with behavioural markers from smartphones for depression diagnosis and monitoring.