Risk Prediction of Metabolic Syndrome for health management cohort:using Random Survival Forests (Preprint)

Abstract

Metabolic syndrome (MetS) is a complex group of metabolic disorder syndromes that are risk factors for diabetes and cardiovascular disease. The main aims of this study were to: (1) to establish a MetS risk prediction model based on a routine health checkup cohort using the RSF algorithm, (2) to compare the prediction performance between the model established by the RSF algorithm and cox regression, and (3) visualize predictive models as a nomogram, which can be adapted for web-based use.

The main aims of this study were to: (1) to establish a MetS risk prediction model based on a routine health checkup cohort using the RSF algorithm, (2) to compare the prediction performance between the model established by the RSF algorithm and cox regression, and (3) visualize predictive models as a nomogram, which can be adapted for web-based use.

The model was based on data from 14,496 subjects aged≥18 years who had a comprehensive health examination and did not have MetS at baseline (2016), with a maximum follow-up of 5 years. A risk prediction model based on the RSF algorithm was developed by applying the physical examination data and compared with the Cox proportional hazards model. The Harrell consistency index (C-index) was used to discriminate the ability of the prediction model, and the Brier score (BS) was used to measure the accuracy of the model prediction results. A nomogram and a web-based application were constructed to vividly demonstrate the prediction model.

During the 5-year follow-up, 593 (4.09%) participants were diagnosed with MetS. In the RSF prediction model, 15 predictors were identified and then based on the Akaike information criterion strategy. we filtered the set of variables with the lowest OOB error rate (OOB=0.130) among the 15 predictors, ultimately containing seven predictors of WC, GLU, HDL-C, BMI, TG, SBP, and Age. In the testing dataset, the C-index scores of the RSF model, the simplified RSF model, and the Cox model were 0.984, 0.978, and 0.875, and the brier scores were 0.046,0.048 and 0.305, respectively, showing that the discriminatory ability and the accuracy of the prediction results of the RSF model were superior compared with the cox model.

We developed a MetS prediction model using the RSF algorithm, which was validated to outperform the cox model, and we constructed a nomogram that can be effectively used for the initial risk prediction of MetS. It is useful for the early identification and prediction of people at high risk of developing MetS.