Abstract
<title>Abstract</title> <p>Background Declining physical fitness and the coexistence of underweight and obesity are growing public-health concerns among university students. Most evidence comes from single cross-sectional surveys that treat repeated student records as independent and rarely build or validate models that could support campus health monitoring. Objective To develop and temporally validate a repeated-measures, surveillance-based model for predicting poor physical fitness among university students, and to characterise the nonlinear contribution of body mass index (BMI) and the potential dual risk associated with underweight and obesity. Methods We analysed routine campus physical fitness surveillance records collected at one Chinese university from 2020 to 2024. Records were linked across years by anonymised student ID. Poor physical fitness was defined as a failing grade under the National Student Physical Health Standard (total score < 60). Within-student correlation was handled using generalized estimating equations (GEE, exchangeable working correlation, robust standard errors) for association analysis, with a random-intercept mixed-effects model as a sensitivity analysis. Nonlinearity of BMI was examined with restricted cubic splines. A pre-test screening model (Model A: sex, age, year of study, testing year, BMI category/continuous) was developed on 2020–2023 data and temporally validated on 2024 data; a partial-surveillance model adding measured fitness components (Model B) was explored. Discrimination, calibration, and clinical usefulness were assessed. Results After cleaning, 107,335 records from 45,191 unique students were analysed (median 2 records per student, IQR 1–3; 74.3% contributed ≥ 2 records). The development set (2020–2023) comprised 82,641 records from 38,070 students and the temporal validation set (2024) comprised 24,694 records from 24,694 students. Poor physical fitness prevalence followed a J-shaped pattern across BMI categories (underweight 8.3%, normal weight 5.6%, overweight 15.2%, obesity 35.8%). In the adjusted GEE model, both extremes carried elevated risk relative to normal weight (underweight OR 1.54, 95% CI 1.43–1.65; overweight OR 2.68, 95% CI 2.51–2.85; obesity OR 7.58, 95% CI 7.03–8.16), and the mixed-effects model gave consistent estimates. BMI showed a significant nonlinear association (P for nonlinearity < 0.001), with fitted risk minimised near a population-level BMI of about 20 kg/m² (lowest-risk range ≈ 19–21 kg/m²). Model A showed moderate discrimination (development AUC 0.755, 95% CI 0.750–0.761; 2024 temporal validation AUC 0.675, 95% CI 0.658–0.691) but over-predicted risk in 2024 (calibration intercept − 1.84, slope 0.68), reflecting a marked between-year drop in prevalence. The partial-surveillance Model B reached high discrimination (2024 AUC 0.945) but is subject to outcome circularity and is reported as exploratory. Conclusions Using five years of student-linked campus surveillance records, a repeated-measures pre-test model identified students at higher risk of poor physical fitness with moderate, honestly reported performance, while a fuller surveillance model achieved high but partly circular discrimination. BMI contributed nonlinearly, and both underweight and obesity were associated with poorer fitness after accounting for within-student correlation. The fitted low-risk BMI range should be read as a population-level surveillance reference, not an individual weight target. Local recalibration and external validation are needed before deployment in campus health management.</p>