Using CGSS2021 data and seven machine learning models with SHAP, the study found that 'Health Impact', 'Age' and 'Depression' were common core predictors of self-rated health in both exercise-participating and non-participating adults, though their ranking and contribution distributions differed, and the cross-sectional single-year design requires further validation.
Key Findings
Methods
Seven machine learning models were built to predict self-rated health (SRH) separately for adults who did and did not participate in physical exercise, using CGSS2021 data.
Data came from the 2021 Chinese General Social Survey (CGSS2021).
Adults were divided into a Participate group and a Non-Participate group.
Seven machine learning models, including Random Forests (RF), were developed to predict SRH status.
The SHAP method was applied to quantify the contribution of individual features to model predictions.
Results
Different top-performing models emerged in the two exercise groups, with small performance differences between the primary models.
In the Participate group, the RF model exhibited a relatively high AUC on the test set.
In the Non-Participate group, the LGBM model exhibited a relatively high AUC on the test set.
The paper states that performance differences between the primary models in the two groups were generally small.
Results
'Health Impact', 'Age' and 'Depression' were common core predictors of SRH across both exercise participation groups, though their importance rankings and contribution distributions were not entirely consistent.
SHAP analysis was used to rank feature contributions in each group.
The rankings and distributions of contributions differed between the Participate and Non-Participate groups.
Results
'Internet Use' and 'Social Friends' showed some auxiliary predictive value for SRH.
These variables were described as having auxiliary rather than core predictive value.
They point to a role for social behaviors in SRH prediction.
Results
Stratification by exercise frequency and sensitivity analyses showed changes in model performance and some feature rankings, but the core predictors remained broadly stable.
Analyses were further stratified by physical exercise frequency.
Sensitivity analyses were also conducted.
Model performance and the ranking of some features changed, while the core predictive factors were broadly stable.
Conclusions
Predicting adult SRH requires multi-level information, including functional status, demographic characteristics, recent emotional experiences and social behaviors.
Key predictors among adults with different levels of exercise participation share commonalities but also show some differences.
The authors suggest the findings may provide a reference for future health research and stratified health promotion.
Discussion
The study's cross-sectional, single-year design limits external validity and practical applicability of the models.
The data were cross-sectional survey data from a single year (2021).
The authors state that external validity and practical applicability require further validation.
What This Means
This study used machine learning to look at which factors best predict how adults in China rate their own health, and whether those factors differ between people who exercise and those who do not. Using 2021 national survey data, the researchers built seven prediction models for each group and used a method called SHAP to see which factors mattered most to each model's predictions.
In both groups, the same three factors stood out: the extent to which health affects daily life ('Health Impact'), age, and depression. How strongly each one contributed differed somewhat between exercisers and non-exercisers. Internet use and having social friends added a smaller amount of predictive information. Different models performed slightly better in each group (Random Forest for exercisers, LGBM for non-exercisers), but the differences were small. The main findings held up when people were grouped by how often they exercised and in sensitivity checks.
This research suggests that how people rate their health reflects a mix of physical function, age, recent emotional state and social life, and that these patterns are broadly similar regardless of exercise participation. Because the data come from a single-year cross-sectional survey, the models show associations only, not causes, and they would need to be tested on other data before being relied on in practice.
Check Your Own Numbers
Upload your bloodwork. We'll cross-reference your results against this study and 4,700 others.
Fan R, Zhang X, Shen J, Sun F. (2026). Machine learning prediction of self-rated health among adults under different physical exercise participation statuses.. Frontiers in public health. https://doi.org/10.3389/fpubh.2026.1882082