Enhancing psychological wellbeing in STEM education through public health strategies across educational transitions

Background Student mental health is a growing public health concern in higher education, and small questionnaire datasets are often highly imbalanced and difficult to model reliably. Objectives This study empirically validates a Transition-Aware Public Health Wellbeing Network (TAPHW-Net) using an openly accessible questionnaire dataset and narrows the STEM-transition domain to the available variables. Methods The source dataset contained 843 responses. Respondents aged 18–22 years were retained, and the analytic sample consisted of 777 students, classified based on the original survey label: No Stress ( n = 38), Eustress ( n = 710), and Distress ( n = 29). 24 predictors were grouped into emotional/psychological, physical/somatic, academic pressure, social and environmental, and demographic branches. To prevent leakage of the targets, the direct recent-stress item was omitted. Median imputation, standardization, and SMOTE were applied in the training folds only. Five-fold out-of-fold evaluation was used to assess Logistic Regression, SVM, Random Forest, XGBoost, MLP, ANN, and TAPHW-Net. Results Logistic Regression had the highest fold-averaged accuracy (0.968 ± 0.019), macro-F1 (0.884 ± 0.048), and macro-AUC (0.996 ± 0.004). TAPHW-Net’s pooled OOF accuracy, macro-F1 score, and macro-AUC were 0.8739, 0.6626, and 0.9523, respectively. It had lower overall accuracy, but it successfully retrieved 78.9% of the No Stress cases and 86.2% of the distress cases, although precision was only 0.385 and 0.397 for these two classes, respectively. Difficulty in the home environment, anxiety, palpitations, work-environment stress, sleep problems, and health problems were important predictors. Conclusion The attention architecture did not outperform simpler models on this small and highly imbalanced tabular dataset. Its main value lies in domain-structured representations, minority-class recall, and transparent mapping of model importance to a hypothetical decision-support template. The findings are limited to the source dataset and require external validation before institutional use or transfer to STEM-specific and educational-transition settings.