Student dropout rates are a major challenge for higher education. Both parents and academic institutions are seeking to reduce this phenomenon by investigating its root causes, as it has economic, social and institutional consequences. Based on an academic dataset enriched by SMOTE balancing, one-hot encoding and scaling, this article explores the application of machine learning (Random Forest, Gradient Boosting) and deep learning (MLP and CNN) techniques to predict this phenomenon and consider possible solutions. Exploring various data, such as grades, absences, failures, study time and family support, the models were compared through these metrics: Accuracy, F1-score, AUC-ROC, PR-curve. The results reveal the effectiveness of Gradient Boosting and Random Forest models (with an F1-score close to 1) over those of Multilayer Perceptron (F1-score = 0.84) and Convolutional Neural Networks (F1-score = 0,82). Analysis of the variables confirms the importance of mid-term marks (G2), absences and previous failures as key predictors. The article provides recommendations, including the parameters to be taken into account in early prediction, and opens up prospects for future work.