Comparative Machine Learning and Explainable AI Framework for Early Mental Health Risk Detection among University Students in Sri Lanka

Abstract

Psychological distress is a concerning scenario among University Students in Sri Lanka, which is compounded by academic demands, financial difficulties, and poor lifestyle patterns. The traditional methods of screening usually cannot detect students “At-Risk” at the early stages and this restricts chances of acting in time. This research proposes an explainable AI-based screening model that allows identifying mental health risks early on in university students. A structured survey of 500 students was conducted as a way of collecting data, including demographic and lifestyle variables, as well as the DASS-21 psychological assessment. A two-step labeling approach that is clinically aligned was implemented and students were initially categorized into two groups “At-Risk” and “Not-At-Risk” based on established DASS-21 cut-offs, and thereafter the “At-risk” students were stratified in terms of their severity, to aid in resources prioritization. Several separate machine learning models were tested and two-step screening strategy was adopted with the help of Logistic Regression as a binary risk predictor and a Decision Tree as a severity predictor. The best binary classification performance was 84 % accuracy with Logistic Regression and ROC-AUC of 0.908 and the best severity stratification performance was 65.3 % accuracy with the Decision Tree. The identified SHAP-based explainability factors are the help-seeking behavior, the regularity of meals, socialization, and the duration of sleep which contribute to the risk of mental health. The proposed system is act as a decisionsupport tool for university counseling services that will be able to identify problems early and intervene in the situation to prevent overuse of available resources.

Ähnliche Arbeiten