Machine learning offers limited screening for poor fitness in Tibet

A cross-sectional study of 5,750 adults aged 20–69 years at three high-altitude locations in Tibet found that interpretable machine-learning models could identify current poor physical fitness with only moderate discrimination. A random forest produced the highest pooled held-out area under the receiver operating characteristic curve, or AUROC, at 0.664 (95% confidence interval, 0.649–0.681). The researchers said the models may help prioritize people for fuller fitness assessment in settings where direct testing is difficult, but they should complement rather than replace such testing and require independent external validation before use.

The analysis drew on data from the Tibet National Physical Fitness Monitoring program. Participants came from Nyingchi at 3,100 metres, Lhasa at 3,658 metres and Nagqu at 4,500 metres; 5,750 of the 6,160 sampled residents were included in the analysis. Poor physical fitness was defined with age-specific total-score cutoffs from the Chinese National Physical Fitness Standard: fewer than 23 points for adults aged 20–39 years, fewer than 18 for those aged 40–59 years and fewer than 15 for those aged 60–69 years. Overall, 2,046 participants, or 35.6%, met the definition. The proportion varied markedly by site, from 27.6% in Lhasa to 46.9% in Nyingchi, showing that the observed prevalence was not uniform across the three locations.

To limit outcome leakage, the investigators prespecified predictors that could be obtained before the scored performance battery: demographic characteristics, exercise behavior, adiposity markers and resting hemodynamic measurements. They compared logistic regression, random forest and XGBoost with internal–external cross-validation by site, using a leave-one-city-out design in which data from each location were assessed with models developed from the other two. Pooled held-out predictions were then evaluated for discrimination, precision-recall, calibration, classification performance at clinically relevant thresholds and decision-curve performance, with 95% confidence intervals obtained through bootstrap resampling. Random forest had the highest pooled held-out discrimination point estimate, followed by XGBoost at 0.652 (95% CI, 0.637–0.667) and logistic regression at 0.646 (95% CI, 0.631–0.661). Differences among the models were small. The random forest’s area under the precision-recall curve was 0.560 (95% CI, 0.537–0.583), and its Brier score was 0.212 (95% CI, 0.208–0.216), while XGBoost achieved a slightly lower Brier score. Calibration of the random forest was moderate, with an intercept of −0.26 and a slope of 0.86.

Performance changed with the threshold used to classify participants. At a prevalence-matched threshold, the random forest achieved a sensitivity of 0.71 and specificity of 0.48. Lowering the fixed threshold to 0.30 raised sensitivity to 0.81 but reduced specificity to 0.33. These results underscore the tool’s screening role: it could flag more people for further assessment, but it would also classify more people without poor fitness as positive. The results were robust in complete-case analysis, under multiple imputation, where the random-forest AUROC was 0.667, and in analyses using a leaner predictor set. An anthropometry-free model performed substantially worse, with an AUROC of 0.55, underscoring the contribution of body-size and adiposity information to the model’s predictions.

Interpretability analysis identified BMI, education level, height, sex, arm skinfold thickness, urbanicity, exercise intensity, waist-to-height ratio and systolic blood pressure as dominant contributors. The ordering of the 12 leading predictors showed substantial agreement with random-forest permutation importance, with a Spearman correlation of 0.78. The authors cautioned that the prominence of body-size variables partly reflects the structure of the Chinese National Physical Fitness Standard, whose anthropometric scoring component contributes to the definition of poor fitness. Overall, the site-wise validation indicated limited but non-trivial transportability across the three Tibetan settings. Because the research was cross-sectional and restricted to the represented adult population, it evaluated identification of existing fitness status rather than changes over time, prospective outcomes or whether screening-driven interventions improve health.

Subject of Research: Machine-learning screening for poor physical fitness among adults at three high-altitude locations in Tibet

Article Title: Interpretable machine learning for identifying poor physical fitness in adults across three high-altitude regions of Tibet: a cross-sectional study

Article References: Chen, T., Zhao, X., Sun, R., Bian, B., & Liang, W. (2026). Interpretable machine learning for identifying poor physical fitness in adults across three high-altitude regions of Tibet: a cross-sectional study. BMC Public Health. https://doi.org/10.1186/s12889-026-29461-z

Image Credits: AI Generated

DOI: 10.1186/s12889-026-29461-z

Keywords: machine learning, physical fitness, high-altitude health, Tibet, random forest, XGBoost, cross-sectional study, screening, adiposity, model interpretability, Interpretable, machine