An Explainable Hybrid Transformer Framework for ADHD Risk Prediction Among University Students Using Psychological Assessment and Self-Reported Text Data

Background: Attention-Deficit/Hyperactivity Disorder (ADHD) can adversely affect university students’ academic functioning, psychological well-being, and adjustment within higher education. Advances in machine learning and natural language processing (NLP) offer opportunities for data-driven approaches to support early identification of elevated ADHD-related symptoms. Problem and Gap: Identifying university students exhibiting elevated ADHD-related symptoms remains challenging when psychological, behavioural, academic, and self-reported information is considered independently. Existing approaches may therefore inadequately exploit complementary information from structured psychological assessments and contextual representations derived from students’ self-reported textual descriptions. Aim: This study develops and evaluates an explainable hybrid transformer-based machine-learning framework to predict elevated ADHD risk by integrating structured psychological and academic features with textual representations. Methodology: A structured-only model was established as a baseline and compared with MiniLM, multilingual transformer, XGBoost–multilingual transformer, BERT, RoBERTa, and DistilBERT hybrid architectures. Performance was evaluated using accuracy, balanced accuracy, precision, recall, F1-score, and ROC-AUC. SHapley Additive exPlanations (SHAP) provided global and individual-level explanations, while McNemar’s test assessed statistical differences between selected model pairs. Results: DistilBERT achieved 71% accuracy, 72% balanced accuracy, 86% precision, 69% recall, 77% F1-score, and 73% ROC-AUC. XGBoost–multilingual transformer attained the highest recall (89%) and F1-score (80%) but only 57% balanced accuracy. McNemar’s tests found no statistically significant differences among the evaluated BERT, RoBERTa, and DistilBERT model pairs. The ROC curve demonstrates that the hybrid multilingual approach offers the best overall trade-off between sensitivity and specificity, supporting the value of combining structured features with multilingual contextual representations for explainable ADHD risk prediction. The SHAP summary plot illustrates that the proposed hybrid model is highly interpretable, revealing that psychological measures, particularly anxiety and depression, are the primary drivers of ADHD risk prediction, while transformer-derived contextual features provide additional predictive information that enhances the model's overall decision-making process.