School dropout has profound and long-term impacts on global development. Machine Learning (ML) techniques have been used to create models to predict school dropouts, supporting school managers, educators, and policymakers to take proactive measures to reduce those rates and their social impact. However, most studies did not account for the imbalance of historical datasets when training those models, leading to over-optimistic performance metrics and poor practical results. In the present work, a novel ANN approach able to deal with imbalanced datasets to predict school dropout is proposed and tested with actual data, reaching the remarkable performance of correctly predicting 90.1% of the students who would dropout and outperforming the benchmark.