Construction and Evaluation of the Prognostic Models

Based on the gene expression values in tumors, the DEGs obtained in method 2.3 were used to construct prognosis-related DEG models using the random survival forest (rsf) method. Random survival forest is obtained by adding survival analysis to random forest analysis, using the bootstrap resampling method to randomly extract N samples from the original dataset. Then, we built a survival trees model to obtain the variable importance measure (VIMP) of each variable. The larger the VIMP value, the stronger the prediction ability. The cumulative hazard function (CHF) was obtained from the mean of each survival tree, which reflects the cumulative probability.²⁵ The data of the DEGs and clinical variables were randomly divided into test and validation groups using the R package, random Forest SRC, to perform the RSF analysis.

We calculated the risk score of CHF based on test set and used the median risk score as the threshold for the high- and low-risk groups. The prognosis-related differential gene set was used as the validation set. The risk score of each sample was obtained by fitting it onto the same model parameters, and the samples were stratified into high- and low-risk groups using the Risk score threshold. We plotted the Kaplan-Meier curve using the “survminer” R package and performed a Log rank test. Time-dependent ROC analyses were conducted using the timeROC package in R. Time-dependent ROC analysis was used to predict the classification efficiency of the RiskScore. Subsequently, we analyzed the classification efficiency of the 1, 3, and 5-year overall survival. A multivariate Cox regression analysis using clinical data was conducted. We used the R package, caret, to conduct the probabilistic calibration analysis and plotted a calibration curve.