DECISION SUPPORT IN IDENTIFYING MENTAL HEALTH RISK PROFILES WITH DEEP NEURAL NETWORKS AND EXPLAINABLE ARTIFICIAL INTELLIGENCE TECHNIQUES

This work applies deep neural networks and Explainable Artificial Intelligence (XAI) techniques to the prediction of clinical levels of anxiety (GAD-7), social phobia (SPIN) and life satisfaction (SWL) from behavioral variables of online gamblers. The data were obtained from a public dataset and processed by a pipeline with stratified partitioning and balancing by oversampling. Three multiclass models were trained using dense layers, ReLU, dropout and Softmax output. The evaluation included accuracy, confounding matrices and PCA projection. For interpretability, LIME and importance by permutation were applied. The models achieved accuracy of 81.9% for GAD-7, 81.7% for SPIN and 74.9% for SWL. The XAI analysis indicates that temporal (Hours, streams) and motivational (whyplay) variables concentrate the predictive signal. As a contribution, the study proposes an interpretable decision support model for initial mental health screening in digital environments.