Counterfactual Explanations and Predictive Modeling of Depression Diagnosis in Brazilian Adults: Exploring Near-Decision-Boundary Cases

This study presents a predictive modeling approach for depression diagnosis among Brazilian adults from the National Health Survey (NHS) 2019, using machine learning models interpreted through counterfactual explanations, an approach within eXplainable Artificial Intelligence (XAI), to enhance interpretability. We applied an adapted CSSE method to generate counterfactuals that respect logical and structural constraints inherent to the survey-based data. The XGBoost model achieved the highest F2-score and Recall for the positive (depression) class, with 0.81 and 0.84, respectively, demonstrating accurate classification of depression cases. The analysis highlighted key attributes whose minimal changes can alter predictions for depression, such as ‘depressed or hopeless feelings’, ‘other mental health disorders’, and ‘sleep medication intake'. These factors reveal model sensitivities and provide insights into decision boundaries, supporting individualized interpretation.