Higher education dropout rates have emerged as a critical global concern, underscoring the urgent need for targeted interventions to enhance student retention and academic success. Addressing the persistent challenge of student attrition in higher education, this study aims to present the Pegaso Student Dropout Prediction (PSTUD) framework, an innovative methodology that leverages Large Language Models (LLMs) in conjunction with established educational and psychological theories to predict university student dropout, and to evaluate its performance in comparison to conventional machine learning approaches. Building upon recent advancements in Machine Learning, the PSTUD framework incorporates key elements from Tinto’s student integration theory, Atkinson’s expectancy-value theory, Hardré and Reeve’s model, and Self-Determination Theory. The framework systematically analyzes various dimensions of the dataset to provide a nuanced understanding of dropout risk factors, while ensuring compliance with ethical standards through Institutional Review Board (IRB) approval and adherence to GDPR regulations. By transforming diverse data types into a unified natural language format, the PSTUD framework captures subtle patterns and interactions that traditional methods might overlook, enhancing predictive accuracy and interpretability. The model’s architecture includes an ensemble of Machine Learning algorithms, such as Random Forest, SVM, and Gradient Boosting, with XGBoost as a meta-learner and integration of Large Language Models (LLMs) has been explored to enhance the prediction of student dropout rates. The transformative potential of LLMs in educational analytics, particularly in predicting and addressing student dropout in higher education by capturing nuanced patterns and providing interpretable insights, offers promising avenues for institutions aiming to improve student retention and success. This integrated approach offers a comprehensive and theoretically grounded method for predicting dropout, providing actionable insights for targeted interventions in higher education. This proposed framework represents a significant advancement in addressing the critical challenge of global university dropout. The findings highlighted that LLMs outperformed baseline Machine Learning models while providing textual analyses of students’ data, showcasing the potential of LLMs as predictive assistants in educational institutions considering data privacy and potential biases within training data.