Beyond Marks: Detecting Hidden Academic Risk Through Psychosocial Proxy Indicators, Machine Learning and Explainable AI in Sri Lankan Government Secondary Schools

Abstract

Background: Academic monitoring in Sri Lankan government secondary schools relies exclusively on term marks, leaving the psychosocial conditions that accumulate into failure entirely invisible. No validated, explainable early warning system exists for this context, and global early warning systems depend on Learning Management System infrastructure unavailable in Sri Lankan government schools. Methods: A 24-item psychosocial questionnaire spanning five theoretically grounded domains — Socioeconomic Burden, Cognitive Load, School Environment, Motivation, and Physical Readiness — was developed through five months of structured field engagement with 12 school principals and 45 classroom teachers, validated by the Professional Psychological Counselors Association of Sri Lanka (PPCA), and approved through three institutional levels. Combined with six-term academic trajectory data from 873 Grade 11 students across 11 government schools in Piliyandala Educational Zone, eight interpretable features were engineered. Four candidate models were evaluated using 10-fold stratified cross-validation; logistic regression was selected. SHAP (SHapley Additive exPlanations) LinearExplainer was applied to produce exact, student-level risk explanations. Results: The full model achieved CV AUC=0.878 (95% CI: 0.858–0.899), a validated 6.1-point gain over the academic-only baseline (AUC=0.818), confirming independent psychosocial contribution. Failure rates rose monotonically from 7.1% to 58.2% across the psychosocial burden spectrum. SHAP analysis identified Cognitive Load and Motivation as the primary actionable risk drivers beyond marks. Critically, 219 students (25.1%) carrying above-median psychosocial burden while currently passing were identified as entirely invisible to conventional marks-only monitoring. Gender fairness evaluation confirmed equitable model performance across four independent statistical tests. Conclusion: The resulting EduRisk prototype delivers SHAP-grounded, student-level risk profiles covering the psychosocial dimensions that marks-only monitoring structurally cannot detect, providing teachers and school authorities with actionable intelligence to intervene before psychosocial burden accumulates into irreversible academic failure. The methodology offers a replicable framework for resource-constrained educational contexts globally.