Leveraging machine learning algorithms and explainable AI for predicting mental health disorder treatment at the workplace

Mental health disorders in the workplace pose a significant global public health challenge, often resulting in reduced productivity. Timely and accurate prediction is essential for facilitating early and effective interventions. In this study, we evaluated six machine learning (ML) algorithms namely logistic regression, random forest, gradient boosting (GB), categorical boosting (CB), support vector machine, and neural network for their ability to predict treatment of mental health disorders using data from the Open Sourcing Mental Health (OSMH) Dataset. To enhance model interpretability and transparency, we incorporated explainable AI techniques, specifically SHapley Additive exPlanations (SHAP), to identify the most influential features driving model predictions. Key predictors included work interference, family history, age, available care options, and help-seeking behavior. The top 10 features identified by the CB and GB models were merged to form the final feature set for model development. Our results demonstrate that ML algorithms can effectively predict mental health outcomes in workplace settings, with the SHAP-enhanced CB and GB models showing the highest clinical relevance and potential for supporting early intervention strategies. The CB and GB models, optimized through feature selection using SHAP, both demonstrated superior performance on the testing data. The CB model achieved an accuracy of 0.8452, precision of 0.8278, recall of 0.9058, and F1 score of 0.8651, while the GB model achieved an accuracy of 0.8254, precision of 0.8357, recall of 0.8478, and F1 score of 0.8417. Both models also yielded the highest AUC scores (0.9101 for CB and 0.9070 for GB) indicating excellent discrimination ability. SHAP analysis revealed that work interference and family history were the most influential features positively impacting the prediction of mental health treatment (mean SHAP values: CB = +0.1724 and + 0.0354; GB = +0.1747 and + 0.0184, respectively), whereas coworkers and seeking help negatively impacted prediction (mean SHAP values: CB = -0.0329 and - 0.0120; GB = -0.0332 and - 0.0111, respectively). These findings highlight the potential of interpretable machine learning models to support targeted workplace mental health interventions. However, the results should be interpreted considering the pre-pandemic 2014 survey context, reliance on self-reported data, and the absence of external validation.