Explainable machine learning for mental health prediction from social media behavior: a nested cross-validation study...

Area of Science:

• Computational psychiatry
• Digital mental health
• Machine learning in healthcare

Background:

• Social media behavior offers potential for early detection of psychological distress.
• Existing predictive models often lack transparency, hindering their use in clinical mental health settings.

Purpose of the Study:

• To develop and evaluate an explainable machine learning (XAI) framework for predicting self-reported depression risk using social media behavioral data.
• To identify key behavioral markers associated with depression risk and assess model transparency and reliability.

Main Methods:

• Utilized a dataset of 481 anonymized social media users.
• Employed a nested 5×5 cross-validation strategy to train and test three supervised learning models.
• Integrated SHAP (SHapley Additive exPlanations) for model explainability and assessed model calibration using reliability curves and Expected Calibration Error (ECE).

Main Results:

• Random Forest model achieved the highest performance (accuracy=84.2%, AUC=0.88) and demonstrated well-calibrated probability estimates.
• Identified significant behavioral markers including screen time, passive scrolling, nighttime usage, and stress-driven engagement.
• SHAP analysis provided stable and consistent feature rankings across multiple random seeds, supporting explanation reliability.

Conclusions:

• The developed XAI framework offers a transparent and interpretable approach for depression risk prediction from social media behavior.
• The findings suggest that interpretable model outputs can inform personalized digital interventions for mental health.
• Future research should focus on larger datasets, multimodal data integration, and clinical validation to enhance generalizability.