Routine physical fitness assessments collected from university students in China contained useful signals for identifying students who would fail the following year’s assessment, according to a retrospective longitudinal prediction study. Researchers developed and tested logistic regression, random forest and XGBoost models on linked annual records from 2023 to 2025. The results showed that the practical value of each model depended on the evaluation metric: logistic regression had the highest discrimination and recall, while random forest achieved the highest accuracy and F1 score and the lowest Brier score.
The study used annual physical fitness assessment records from one Chinese university. The model-development dataset contained 11,430 linked student-year pairs spanning 2023–2024, and the out-of-time test dataset contained 11,427 linked pairs spanning 2024–2025. This temporal separation tested whether models fitted to earlier records could predict outcomes in a later cohort.
Predictors came from the preceding year’s assessment and included body mass index, vital capacity, 50-meter sprint time, standing long-jump distance, sit-and-reach distance, endurance run time and sex-specific muscular strength. The investigators compared the three algorithms using discrimination, calibration and classification metrics at a prespecified probability threshold of 0.50. They also applied SHapley Additive exPlanations, or SHAP, to the fitted XGBoost model to characterize the contribution of each predictor to its output.
No single algorithm led on every measure. Logistic regression achieved the highest area under the receiver operating characteristic curve, with an ROC-AUC of 0.8167, and the highest precision-recall area under the curve, with a PR-AUC of 0.3227. Its recall was 0.7569, the highest among the three models, indicating that it detected the largest proportion of subsequent-year failures at the chosen threshold.
Random forest recorded the highest accuracy, at 0.9320, and F1 score, at 0.3573, as well as the lowest Brier score, at 0.0578. Its recall was 0.3000, however, and its calibration slope was 0.4985. XGBoost produced an ROC-AUC of 0.7935 and a recall of 0.5458.
SHAP analysis identified body mass index, endurance run time, standing long-jump distance and vital capacity as the largest overall contributors to XGBoost’s model output. The analysis quantified feature contributions within the fitted algorithm, but it did not establish that changing any individual measurement would cause a student to pass a later assessment or that the same feature was most influential for every student. The model’s performance was also sensitive to how failure was defined: sensitivity analyses found lower predictive performance when subsequent-year failure was based on percentile-based thresholds rather than the absolute total-score cutoff of 60.
The findings therefore depend in part on the outcome definition applied to the annual assessment. Several limitations constrain the conclusions. Although the predictors preceded the outcomes they were used to predict, both the measurements and the subsequent-year total fitness score were derived from the same physical fitness assessment framework.
The authors noted that persistence of the underlying fitness construct and conceptual overlap between predictors and outcomes may have contributed to the observed performance. In addition, the out-of-time evaluation occurred within a single university and did not constitute independent external validation. The researchers said that external validation, local calibration, threshold assessment and prospective evaluation would be required before the models could be used for routine student monitoring or intervention.
Subject of Research: Longitudinal machine-learning prediction of annual physical fitness assessment failure among university students
Article Title: Predicting subsequent-year physical fitness assessment failure in university students using longitudinal data and explainable machine learning
Article References: Luo, L., Huang, F., & Li, B. (2026). Predicting subsequent-year physical fitness assessment failure in university students using longitudinal data and explainable machine learning. BMC Public Health. https://doi.org/10.1186/s12889-026-29742-7
Image Credits: AI Generated
DOI: 10.1186/s12889-026-29742-7
Keywords: physical fitness, university students, machine learning, longitudinal data, risk prediction, XGBoost, random forest, logistic regression, SHAP analysis, out-of-time validation, Predicting, subsequent-year
News Source: Teresa Odom. (October 26, 2026). Machine learning predicts next-year fitness assessment failure in university students. Scienmag.
