MINDCARE: AN INTERPRETABLE MULTIMODAL FRAMEWORK FOR EARLY STUDENT MENTAL- WELLNESS RISK ASSESSMENT USING NLP AND BAYESIAN INFERENCE

Student mental-wellness assessment involves heterogeneous evidence, includ- ing self-report, journal text, and optional behavioural indicators. We present MindCare, a lightweight evidence-fusion framework that combines a 12-question instrument incorporating DASS-21-derived and CHU9D-derived items, rule- assisted linguistic features, and normalized sleep/activity indicators within an in- terpretable Bayesian Network. Journal text is processed using phrase matching, domain-specific keywords, VADER sentiment, and TextBlob polarity/subjectivity. The network produces posterior estimates for anxiety, depression, and stress. Evaluation comprises functional testing on 20 controlled cases, external evaluation of four NLP-derived features with conventional classifiers, and Bayesian sensitiv- ity and evidence-source ablation analyses. The internal functional test achieved 90.0% accuracy, whereas the external seven-class dataset yielded macro F1 scores of 21.61%, 21.01%, and 26.56% for Gaussian Naive Bayes, Logistic Regression, and Random Forest, respectively. Sensitivity and ablation experiments showed directionally consistent evidence responses. These results establish technical fea- sibility and interpretability of the implemented evidence-fusion mechanism, but do not establish clinical validity or population-level generalization because the in- ternal test is small, external labels do not match the Bayesian outputs, and CPTs are predefined rather than empirically calibrated.