Systematic Review and Meta-Analysis of Explainable Machine Learning Models for Clinical Depression Detection

Area of Science:

• Computational psychiatry and machine learning applications in mental health.
• Systematic review and meta-analysis of supervised learning algorithms for clinical diagnostics.

Background:

• Depression is a prevalent mental disorder requiring early detection for effective psychotherapy outcomes.
• Supervised algorithms (SVM, Random Forest, XGBoost, GCN) are increasingly explored for clinical depression detection using real-world data.

Purpose of the Study:

• To evaluate the accuracy, interpretability, and generalizability of supervised algorithms for depression detection.
• To assess the impact of data sources and interpretability methods on algorithmic performance in clinical settings.

Main Methods:

• Systematic review and meta-analysis of 20 studies (2014-2025) adhering to PRISMA guidelines.
• Analysis of F1-Score, AUC-ROC, SHAP/LIME interpretability, and cross-validation strategies.
• Statistical analysis using ANOVA and Pearson correlations to assess performance and relationships.

Main Results:

• XGBoost showed the highest average performance (F1: 0.86, AUC: 0.84), but differences between algorithms were not significant.
• SHAP was the dominant interpretability method; combined SHAP+LIME correlated with higher F1-Scores.
• Clinical surveys and EHR data yielded stable results, while neurophysiological data showed high estimates but limited representation; significant heterogeneity and publication bias for AUC were observed.

Conclusions:

• Algorithmic performance in depression detection is more influenced by data quality, context, and interpretability than the specific model.
• Explainable AI approaches provide practical value for personalized and collaborative clinical decision-making in mental health.
• Findings challenge claims of inherent algorithmic superiority, emphasizing a holistic approach to model selection and evaluation.