Exploring Machine-Learning Prediction of Self-Reported Academic Burnout

Abstract Academic burnout is a concern in higher education, but predicting students’ self-reported experiences from small survey datasets presents methodological challenges. This exploratory study examined whether academic and lifestyle survey responses could distinguish three categories of students’ responses to a single item asking whether they believed they were currently experiencing academic burnout: “Yes,” “Not sure,” and “No.” The analysis used 101 survey responses and compared eight classification approaches across three feature sets. Model evaluation used repeated stratified five-fold cross-validation, with hyperparameter selection conducted within the training data. The highest mean macro-F1 score was 0.430 (SD = 0.112), achieved by the k-nearest neighbours classifier using the core feature set. Its mean accuracy was 0.485 (SD = 0.116), compared with 0.436 for the majority-class baseline. Overall performance was modest and varied across validation folds. Class-specific results showed that the model had particular difficulty identifying respondents who answered “No.” Because the outcome was a single self-reported item rather than a validated burnout measure, the results should be interpreted as exploratory evidence of classification feasibility, not as a diagnostic or screening tool. The findings highlight the importance of cautious interpretation, transparent validation, and larger, independently evaluated datasets in research on student well-being.