Low-dimensional burnout risk classification in engineering students: the role of demand-recovery interaction features

Student burnout is a rapidly growing concern, with prior research linking its development to a wide range of psychological, behavioral, and contextual factors. Machine learning approaches have recently shown strong predictive performance for identifying burnout, but typically rely on high-dimensional datasets impractical for real-world use. Whether burnout can be reliably identified from a minimal set of variables remains underexplored. This study addresses that gap using a low-dimensional predictive framework in undergraduate engineering students. Grounded in Conservation of Resources theory, three classifiers (Logistic Regression, Random Forest, Gradient Boosting) were evaluated across three feature configurations (Raw: all nine unengineered base predictors, including academic, demographic, sleep, and physical activity variables; Engineered: four demand–recovery interaction features; Full: Raw combined with Engineered) to test whether burnout reflects academic load alone or an imbalance between demand and recovery capacity. The Full configuration performed best: Logistic Regression and Random Forest exceeded a matched permutation-based null (Global MCC = 0.173 and 0.190, respectively, p < 0.05), while Gradient Boosting did not reach significance (MCC = 0.126, p = 0.066). Across 50 alternative fold splits, however, mean Global MCC for the Full configuration was lower for each classifier (Logistic Regression 0.09, Random Forest 0.16, Gradient Boosting 0.11), indicating that the primary-split estimates are upper-range. In addition, adding engineered features to Raw predictors did not produce a statistically distinguishable improvement in Global MCC for any classifier. Moderate Risk classification remained difficult across all classifiers and configurations, with most confidence intervals crossing zero. Permutation importance identified Workload-Sleep Strain as the most influential predictor in the Random Forest model, suggesting the engineered features capture meaningful behavioral patterns. However, because these variables are deterministic transformations of the raw predictors, their incremental contribution could not be confirmed statistically. These results offer preliminary evidence that a small set of sleep, activity, workload, and engineered variables carries predictive signal for burnout risk above chance.