Predicting O-Level Student Dropout Risk in Southwestern Uganda: A Context-Aware Machine-Learning Early-Warning Model

Abstract

Student dropout remains a major obstacle to equitable secondary education, yet many schools in low-resource settings identify risk only after prolonged absenteeism or disengagement. This article develops and evaluates a context-aware predictive analytics model for early identification of dropout risk among O'Level learners in selected private mixed secondary schools in Mbarara and Isingiro districts, Southwestern Uganda. Guided by design science research, the study analysed complete records from 331 learners sampled from an accessible population of 1,919 across six schools. Academic, socioeconomic, demographic, institutional and psychosocial indicators were encoded and standardised before a stratified 70:30 train-test split (232 training and 99 test records) and five-fold cross-validation. Logistic Regression, Random Forest, Support Vector Machine (SVM) and K-Nearest Neighbours (KNN) were compared, with F1-score prioritised because an early-warning tool must balance detection of at-risk learners against false alerts. SVM produced the strongest reported F1-score (0.633), followed by Logistic Regression (0.600), Random Forest (0.511) and KNN (0.468). Chronic absenteeism, financial hardship and delayed fees, long travel distances, weak examination confidence, bullying, limited parental engagement and psychological stress emerged as salient risk indicators. The selected SVM classifier was integrated into a Python/Qt desktop application supporting individual and batch prediction, risk visualisation and exportable reports. The findings show the feasibility of locally grounded predictive support, but the moderate F1-score, small purposive school frame, researcher-constructed risk labels and absence of external prospective validation preclude autonomous high-stakes use. The model should therefore operate as a human-supervised screening aid accompanied by counselling, welfare and academic support, with wider validation, calibration, fairness audits and transparent data-governance procedures before scale-up.