A Hybrid Deep Learning Model for Early Prediction of Student Academic Performance in Higher Education

Abstract

Early identification of students who may experience academic difficulty can support timely, targeted intervention in higher education. This study evaluates a hybrid prediction framework that combines a support vector machine (SVM) with a polynomial kernel, a long short-term memory (LSTM) component with two hidden units, and a decision-tree component. The study protocol uses cohort-wise 70:15:15 train-validation-test partitioning so that students from different semester, college, course, and course-type cohorts are not mixed across partitions. Hyperparameters were selected using random search, and the component baselines were evaluated under the same protocol. The finalized prediction file supplied for this study contains 4,800 student-level observations, three performance classes (0, 1, and 2), and predictions from SVM, LSTM, decision tree, and the proposed hybrid system. On this 4,800-record prediction set, the hybrid achieved 94.10% accuracy (95% bootstrap CI: 93.44–94.75%), balanced accuracy of 94.05%, macro-F1 of 93.57%, Cohen's kappa of 0.906, and Matthews correlation coefficient of 0.906. The confusion matrix contains 1,371, 2,245, and 901 correct predictions for classes 0, 1, and 2, respectively, with 283 total errors. The supplied component predictions yielded accuracies of 82.94% (SVM), 85.04% (LSTM), and 79.19% (decision tree), providing an empirical comparative reference. A critical data-audit finding is that the supplied CSV does not contain cohort/year identifiers, model probabilities, or a trained-model checkpoint; consequently, native SHAP values and independent construction of year-specific results cannot be claimed from this file alone. This distinction is made explicitly to avoid overstating evidence.