2.6. Assessment of the Clinical-Molecular Integrated Prognostic Model

First, according to the distribution of the PDS of the corresponding molecular factors identified, patients in the discovery set were divided into different groups. The grouping was based on the highest degree of differentiation of survival curves. These findings could provide a direct observation of the relevance of the identified molecular prognostic factors and survival.

Second, based on the clinical prognostic factors and identified molecular prognostic factors, one clinical prognostic model, one molecular prognostic model, and one clinical-molecular integrated prognostic model were constructed on the discovery set. Internal validation through bootstrapping with 200 iterations was used to assess the discrimination performance of these models on the discovery set. External validation through stratified bootstrapping with 200 iterations was used to assess the discrimination performance of these models on the validation set. As the mean survival time of patients with metastasis was shorter than that of patients without metastasis, the performance of the prognostic model might have been affected. Therefore, models for nonmetastatic patients were built with the same prognostic factors and compared with the same assessment.

Finally, to compare the prognostic performance of directly using gene expression data and using converted PDS in this study, genes involved in the pathways were combined with clinical prognostic factors in the clinical-molecular integrated prognostic model. Comparisons between the gene-based integrated prognostic model and pathway-based integrated prognostic model were conducted.

The constructed integrated prognostic model had a potential problem of overfitting as it contains multiple covariates. The bias-corrected Harrell's C-index which overcomes the problem of overfitting was chosen to evaluate the overall discriminative performance of the models in internal validation [28]. The origin Harrell's C-index was used in external validation of the overall discriminative performance. Uno's C-index, which is free of censoring, was chosen to evaluate the discriminative performance of the models at the 3-year time point [29]. A two-sided Wilcoxon signed-rank test was used to compare the 200 C-indexes generated from the 200 iterations of the bootstrapping procedure to quantify the discriminative difference of the C-index between different models.