JMIR Publications

ABSTRACT

Background:

The robust pervasive device-based existing systems to detect depression developed in recent years requiring data collected over a long period of time may not be effective in cases where early detection is crucial. In addition, due to the requirement of running systems in the background for prolonged periods, the existing promising systems may not be resource efficient. As a result, these systems can be infeasible in low-resource settings.

Objective:

Our main objective was to develop a minimal system to identify depression that works on data retrieved in the fastest possible time. Another objective of this study was to explain the machine learning (ML) models which performed best in identifying depression.

Methods:

We developed a tool that retrieves the past 7 days’ app usage data in a second (mean=0.31 second, SD=1.10 second) which is computationally cheaper. In our study, 100 students from Bangladesh participated and our tool collected their app usage data and responses to the Patient Health Questionnaire-9 (PHQ-9) scale. To identify the depressed and non-depressed participants, we developed a diverse set of ML models including linear, tree-based, and neural network-based models. We selected the features by the Stable approach along with the 3 main types of feature selection (FS) approaches: Filter, Wrapper, and Embedded. We developed and validated the models using the Leave One Participant Out Cross Validation (LOPOCV) method. Additionally, we presented the explanations of the best ML models through the SHapley Additive exPlanations (SHAP).

Results:

Among the 100 participants, 51% participants were depressed and 49% participants were non-depressed. Leveraging only the app usage data retrieved in a second, our Light GBM model using the Stable approach selected features identified 82.4% depressed correctly (precision=75%, F1 score=78.5%). Moreover, using around 5 features selected by the all-relevant FS approach Boruta in each iteration of LOPOCV, we presented a parsimonious Stacking model which had a maximum precision of 77.4% and a balanced accuracy of 77.95%. SHAP analysis on our best models presented different app usage behavioral markers that have a relation with depression. For instance, we found that non-depressed students’ spending time on Education apps is higher on weekdays while depressed students used a higher number of Photo & Video apps and also had a higher deviation in using Photo & Video apps over the day of the weekend.

Conclusions:

Due to the faster and minimalistic nature, our system may make a worthwhile contribution to identifying depression in underdeveloped and developing regions. In addition, our discussion about the implication of findings can facilitate the development of resource insensitive systems, and also in better understanding the depressed students and taking steps in intervention.