Withdrawal from foundational programming courses is a considerable hurdle for higher education, impeding the pipeline of skilled professionals and diminishing institutional effectiveness. To address this, the timely detection of vulnerable students is paramount for deploying targeted support. This research details the development and assessment of a machine learning algorithm designed to identify at-risk students within the first four weeks of instruction, using data from 207 students at a Colombian university. Employing the CRISP-DM framework, the analytical process involved testing seven classification algorithms. This was enhanced by data balancing with the Resample technique, feature selection via the ELI5 library, and hyperparameter tuning using Grid Search. The findings revealed that the Resample technique significantly boosted model efficacy. The highest level of performance, indicated by a peak F1 score of 0.87, was obtained using the Support Vector Classification (SVC) algorithm. This favorable outcome is directly attributable to the model’s capacity for effectively managing the inherent trade-off between the precision and recall metrics. The model’s strong classification capability was confirmed in a pilot implementation, where it achieved Area Under the Curve scores surpassing 0.90 for all risk classifications. The study concludes that dependable, early prediction of attrition risk is feasible even with constrained data, provided that academic performance is integrated with motivational and demographic information. This research establishes a robust groundwork for future investigations and the creation of proactive intervention systems to enhance student persistence in vital academic programs.