Machine learning offers limited screening for poor fitness in Tibet
A study of 5,750 adults aged 20–69 at three high-altitude locations in Tibet found that machine-learning models could identify people with poor physical fitness, but their accuracy was limited. These models use computer programs to find patterns in data. A random forest model performed best when tested on data it had not been trained on, with an area under the receiver operating characteristic curve (AUROC) of 0.664 (95% confidence interval, 0.649–0.681). The researchers said these models might help identify people who need more complete fitness tests, especially in places where direct testing is difficult. However, the models should support, not replace, these tests. They must also be tested independently in other groups before they are used.
The researchers used 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 above sea level. Of the 6,160 residents originally sampled, 5,750 were included in the analysis. The researchers defined poor physical fitness using age-based score limits from the Chinese National Physical Fitness Standard. Adults aged 20–39 were considered to have poor fitness if they scored below 23 points. For those aged 40–59, the limit was below 18 points, and for those aged 60–69, it was below 15 points. In total, 2,046 participants, or 35.6%, met this definition. The percentage differed greatly between locations: 27.6% in Lhasa and 46.9% in Nyingchi. This showed that poor fitness was not equally common across the three places.
To avoid giving the models information that would reveal the fitness test results in advance, the researchers chose predictors that could be measured before the fitness tests were scored. These included personal and demographic details, exercise habits, measures of body fat, and measurements of blood flow and pressure while resting. They compared three methods: logistic regression, random forest, and XGBoost. They tested the models using a method that took turns leaving out one location. Each model was trained using data from two locations and tested on data from the third. They then combined the test predictions and assessed how well the models separated people with poor fitness from those without it, how accurately they identified positive cases, how closely their predicted risks matched actual results, how well they classified people at useful decision thresholds, and whether using them might help guide decisions. They calculated 95% confidence intervals by repeatedly resampling the data using a bootstrap method. The random forest model had the highest overall AUROC, followed by XGBoost at 0.652 (95% CI, 0.637–0.667) and logistic regression at 0.646 (95% CI, 0.631–0.661). The differences between the models were small. The random forest had an area under the precision-recall curve of 0.560 (95% CI, 0.537–0.583) and a Brier score of 0.212 (95% CI, 0.208–0.216). The Brier score measures how close predicted risks are to actual outcomes, with lower scores generally indicating better predictions. XGBoost had a slightly lower, and therefore better, Brier score. The random forest's calibration, or how closely its predicted risks matched the observed results, was moderate, with an intercept of −0.26 and a slope of 0.86.
The models performed differently depending on the cutoff used to label someone as having poor fitness. At a cutoff chosen to match the observed rate of poor fitness, the random forest correctly identified 71% of people with poor fitness (sensitivity of 0.71) and correctly identified 48% of people without poor fitness (specificity of 0.48). Lowering the cutoff to a fixed value of 0.30 increased sensitivity to 0.81 but reduced specificity to 0.33. This means the model could identify more people who needed further testing, but it would also incorrectly flag more people who did not have poor fitness. The results remained similar when the researchers analyzed only participants with complete data, used a method called multiple imputation to fill in missing values (the random forest AUROC was 0.667), and tested a model with fewer predictors. A model that did not use body measurements performed much worse, with an AUROC of 0.55. This suggests that information about body size and body fat played an important role in the model's predictions.
The researchers found that body mass index (BMI), education level, height, sex, the thickness of the skinfold on the arm, whether a person lived in an urban or rural area, exercise intensity, waist-to-height ratio, and systolic blood pressure were among the most important factors in the model. The ranking of the 12 leading predictors closely matched the results from a separate random forest test that measured how much each predictor mattered. The Spearman correlation between the rankings was 0.78. However, the authors warned that body-size measurements ranked highly partly because the Chinese National Physical Fitness Standard itself includes body measurements when calculating fitness scores. Overall, testing each location separately showed that the models could be used across the three Tibetan settings to a limited, but meaningful, extent. The study collected information at one point in time and included only the adult population represented in the data. It therefore assessed whether the models could identify people's existing fitness levels, not how fitness changed over time, what health outcomes might occur later, or whether interventions based on screening would improve health.
Subject of Research: Using machine learning to screen 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, body fat, model interpretability
News Source: Teresa Odom. (October 26, 2026). Machine learning offers limited screening for poor fitness in Tibet. Scienmag.