Development and visualization of a risk prediction model for metabolic syndrome: a longitudinal cohort study based on health check-up data in China.

Aim: Our study aimed to construct a practical risk prediction model for metabolic syndrome (MetS) based on the longitudinal health check-up data, considering both the baseline level of physical examination indicators and their annual average cumulative exposure, and to provide some theoretical basis for the health management of Mets.

Methods: The prediction model was constructed in male and female cohorts, separately. The shared set of predictive variables screened out from 49 important physical examination indicators by the univariate Cox model, Lasso-Cox model and the RSF algorithm collectively was further screened by Cox stepwise regression method. The screened predictors were used to construct prediction model by the Cox proportional hazards regression model and RSF model, respectively. Subsequently, the better method would be selected to develop final MetS predictive model according to comprehensive comparison and evaluation. Finally, the optimal model was validated internally and externally by the time-dependent ROC curve (tdROC) and concordance indexes (C-indexes). The constructed predictive model was converted to a web-based prediction calculator using the "shiny" package of the R4.2.1 software.

Results: A total of 15 predictors were screened in the male cohort and 9 predictors in the female cohort. In both male and female cohorts, the prediction error curve of the RSF model was consistently lower than that of the Cox proportional hazards regression model, and the integrated Brier score (IBS) of the RSF model was smaller, therefore, the RSF model was used to develop the final prediction model. Internal validation of the RSF model showed that the area under the curve (AUC) of tdROC for 1 year, 3 years and 5 years in the male cohort were 0.979, 0.991, and 0.983, and AUCs in the female cohort were 0.959, 0.975, and 0.978, respectively, the C-indexes calculated by 500 bootstraps of the male and female cohort RSF models are above 0.7. The external validation also showed that the model has good predictive ability.

Conclusion: The risk predictive model for MetS constructed by RSF in this study is more stable and reliable than Cox proportional hazards regression model, and the model based on multiple screening of routine physical examination indicators has performed well in both internal and external data, and has certain clinical application value.