A Fast and Minimal System to Identify Depression Using Smartphones: Explainable Machine Learning-Based Approach

Area of Science:

• Digital Phenotyping
• Machine Learning in Mental Health
• Computational Psychiatry

Background:

• Existing depression detection systems require extensive data collection over time, limiting their effectiveness for early intervention.
• Current methods can be resource-intensive and infeasible in low-resource environments.

Purpose of the Study:

• To create a minimalistic system for rapid depression identification using minimal data.
• To identify and explain the most effective machine learning models for depression detection.

Main Methods:

• Developed a tool to collect 7-day app usage data within 1 second from 100 Bangladeshi students.
• Employed diverse machine learning models (linear, tree-based, neural networks) with feature selection (filter, wrapper, embedded) and nested cross-validation.
• Utilized Shapley Additive Explanations (SHAP) for model interpretability.

Main Results:

• A light gradient boosting machine model achieved 82.4% accuracy in identifying depressed students using 1-second app usage data.
• A stacking model with ~5 Boruta-selected features reached 77.4% precision and 77.9% balanced accuracy.
• Diurnal app usage patterns were more indicative of depression than aggregated data; SHAP analysis revealed specific behavioral markers associated with depression.

Conclusions:

• The fast, minimalistic system shows promise for depression detection in underdeveloped and developing regions.
• Findings can guide the development of less resource-intensive systems for understanding and intervening in student depression.