Hidden low wellbeing among high-achieving students: model-based support-cost patterns revealed by explainable machine learning.

Introduction: High achievement is often treated as evidence of positive adjustment, yet some relatively high-achieving students continue to perform well while reporting low subjective wellbeing. Guided by Self-Determination Theory and Expectancy-Value Theory, this study examined whether relatively high-achieving students with low and high life satisfaction could be distinguished by support-related and cost-related predictive signals. Methods: PISA 2022 data from 15 education systems were analyzed. Relatively high-achieving students were defined as those in the top 25% of mathematics performance within their education system. Life satisfaction scores of 0 to 4 indicated low life satisfaction, whereas scores of 7 to 10 indicated high life satisfaction. The final sample included 24,947 students, comprising 20,862 students with high achievement and high life satisfaction and 4,085 students with high achievement and low life satisfaction. Eleven machine-learning classifiers were compared, and the selected Gradient Boosting Machine model was interpreted using SHapley Additive exPlanations. Results: The final 12-feature model achieved a weighted validation AUC of 0.810 and an unweighted validation AUC of 0.828. School belonging was the strongest differentiating signal. Higher emotion regulation, family support, relationship quality at school, perceived safety, creative family environment, and self-directed learning efficacy generally shifted predictions toward high life satisfaction. Mathematics anxiety, bullying victimization, and self-directed learning difficulties shifted predictions toward low life satisfaction. SHAP interaction analyses indicated that low-life-satisfaction classification relied more strongly on combinations of support-related and cost-related signals centered on school belonging. Discussion: Hidden low wellbeing among relatively high-achieving students was distinguished by a life-satisfaction-based predictive profile involving relational, emotional, family, peer, and learning-experience signals. The findings may support group-level monitoring and future theory development, but they should not be interpreted as causal evidence or used as an individual diagnostic tool.