SHAP-ALE: A Novel Approach to Explainability in Mental Health Prediction using RFR
TL;DR
This study highlights the contribution of SHAP-ALE as a hybrid explainability framework that integrates local and global interpretability, enabling a more comprehensive understanding of feature interactions and non-linear effects beyond the capabilities of individual methods.
Abstract
The prediction of mental health disorders, such as depression, has become increasingly crucial as global mental health concerns continue to rise. In this study, the predictive task specifically focuses on estimating the prevalence of depressive disorders as the primary target variable, while other mental health conditions such as anxiety and schizophrenia are treated as explanatory features. While machine learning (ML) models, like Random Forest Regressor (RFR), offer high accuracy in predictions, their interpretability remains a challenge. This research introduces SHAP-ALE, an innovative hybrid explainability framework that integrates SHapley Additive exPlanations (SHAP) and Accumulated Local Effects (ALE) to address this gap. SHAP provides both global and local insights into feature contributions, while ALE visualizes feature-target relationships, mitigating bias caused by feature correlations. Using a dataset comprising various mental health disorders and demographic factors, the dataset used in this study was obtained from Kaggle and consists of 6,421 records covering multiple mental health disorder indicators and demographic attributes across different regions and years. RFR model demonstrated robust predictive performance with an R² score of 0.9984 and a Mean Squared Error (MSE) of 0.0016. SHAP analysis revealed that features such as schizophrenia and anxiety disorders significantly influenced predictions, while ALE identified nonlinear relationships between these features and depression prevalence. The combined insights from SHAP and ALE enhance the interpretability of the model, enabling better understanding of the complex factors underlying mental health disorders. This study highlights the contribution of SHAP-ALE as a hybrid explainability framework that integrates local and global interpretability, enabling a more comprehensive understanding of feature interactions and non-linear effects beyond the capabilities of individual methods.