Abstract
<title>Abstract</title> <p>Academic early warning systems are essential for shifting higher education governance from delayed remediation to proactive, data-driven intervention. However, existing warning mechanisms often depend on fragmented academic records or single-source indicators, limiting their ability to identify students at different levels of academic risk in time. This study develops and evaluates an intelligent multi-class academic early warning framework using a real higher education dataset containing 1,194 student samples and 31 multi-modal features. The feature space integrates academic indicators, socio-economic and demographic variables, behavioral and lifestyle characteristics, and psychological, health, and living-state information. After missing-value imputation, categorical encoding, and numerical standardization, exploratory data analysis was used to examine grade distribution and feature interactions, while t-distributed stochastic neighbor embedding was applied to verify the separability of academic risk groups in a low-dimensional manifold space. A One-vs-Rest multi-class classification architecture based on advanced ensemble tree models was then constructed to classify students into low-risk, medium-risk, and high-risk categories. The empirical results show that previous semester performance is the strongest positive predictor of current academic achievement, while attendance, study time, and daily behavioral patterns provide additional discriminative information. The t-SNE projection reveals clear clustering tendencies for high-risk students, and the proposed classifier achieves strong test-set discrimination, high prediction confidence, a macro-averaged AUC above 0.95, robust precision-recall performance under class imbalance, and stable convergence without severe overfitting. These findings demonstrate that multi-modal learning analytics can provide reliable decision support for timely academic counseling, personalized tutoring, and student retention management.</p>