Journal of Artificial Intelligence, Applications and Innovations 3060-7124 2 Journal of Artificial Intelligence, Applications and Innovations 2025 04 01 STEM-Net: An Explainable Multi-Task Hybrid Architecture Combining Tabular Transformers, Gradient-Boosted Trees and Ensemble Learning for the Simultaneous Prediction of Eight Psychological and Academic Problems in High-School Students STEM-Net: An Explainable Multi-Task Hybrid Architecture Combining Tabular Transformers, Gradient-Boosted Trees and Ensemble Learning for the Simultaneous Prediction of Eight Psychological and Academic Problems in High-School Students 27 47 EN Journal Article 2026 06 28 <p>Psychological and academic problems in students rarely occur in isolation: depression, anxiety, low self-esteem and addiction proneness tend to appear together, and modelling each of them separately wastes the signal they share. This study introduces the STEM-Net framework. Data were collected from 1,713 high-school students using eight validated questionnaires; after synthetic minority oversampling (SMOTE) the set grew to 5,139 records, and labels were coded at several risk levels based on the official cut-off points of each instrument. This multi-task hybrid architecture predicts all eight problems simultaneously, each at the risk levels defined by its own instrument. The base layer contains five heterogeneous learners in three branches: a multi-task tabular transformer whose attention body is shared across the eight targets, gradient-boosted tree models trained per target, and a pre-trained network. Before fine-tuning, the body of the deep branch is pre-trained on roughly eighty-eight thousand public respondents to two psychometric instruments, and the outputs of the five learners are combined through an automatic choice between stacking and simple averaging. To prevent item-to-score leakage, the items of the target scale are removed from the feature space of that task. On the test set STEM-Net reaches a macro-average accuracy of 0.9277, an F1 of 0.9034 and a ROC-AUC of 0.9844. A four-method explainability analysis shows that in every one of the eight tasks the most informative source is another one of the scales, and that addiction proneness and suicidal ideation are mutually predictive. Building on these features, a web-based screening system with a short questionnaire was designed that reports, for each student, the predicted risk level together with the features that most influenced it.</p> https://journalaiai.com/index.php/aiai/article/download/111/42