This article aimed to develop a classification model for predicting student academic performance by representing outputs into various multi-level representations with better accuracy using supervised learning algorithms. The prediction model developed is an integral part of the intelligent academic recommendation system. The method used is a backpropagation neural network with a multi-layer perceptron architecture. The activation function uses the Tanh and SoftMax functions in each hidden layer and output layer, respectively. We train models on real-world student academic datasets using the K-fold cross-validation and the confusion matrix techniques to measure accuracy, precision, and recall results. The tested result shows that the effectiveness of the models and methods used produces the best accuracy of 80.95%, 83.20% precision, and 83.53% recall with a cross-entropy loss rate of 0.12. The Area Under Curve (AUC) Receiver Operating Characteristics (ROC) curve produces a value of 0.89, which is classified as a good result. Based on the test result, the model can be applied to the student academic early warning mechanism in the intelligent academic recommendation system to support the education process in higher education.