No models have been developed to predict the survival probability for women with primary vaginal cancer (VC) due to VC’s extreme rareness. We aimed to develop and validate models to predict the overall survival (OS) and cancer-specific survival (CSS) of VC patients.
A population-based multicenter retrospective cohort study was carried out using the 2004–2018 Surveillance, Epidemiology, and End Results Program database in the United States. The final multivariate Cox model was identified using the Brier score and Harrell’s C concordance statistic (C-statistic). The decision curve, calibration plot, and area under the time-dependent receiver operating characteristic curve (AUC) were used to evaluate model prediction performance. Multiple imputation followed by bootstrap was performed. Bootstrap validation covered the entire statistic procedure from model selection to baseline survival and coefficient calculation. Nomograms predicting OS and CSS were generated.
Of the 2,417 eligible patients, 1,692 and 725 were randomly allocated to the training and validation cohorts. The median age (Interquartile range) was 66 (56–78) and 65 (55–76) for the two cohorts, respectively. Our models had larger net benefits in predicting the survival of VC patients than the American Joint Committee on Cancer stage, presenting great discrimination ability and excellent agreement between the expected and observed events. The performance metrics of our models were calculated in three cohorts: the training cohort, complete cases of the validation cohort, and the imputed validation cohort. For the OS model in the three cohorts, the C-statistics were 0.761, 0.752, and 0.743. The slopes of the calibration plots were 1.017, 1.005, and 0.959. The 3- and 5-year AUCs were 0.795 and 0.810, 0.768 and 0.771, and 0.770 and 0.767, respectively. For the CSS model in the three cohorts, the C-statistics were 0.775, 0.758, and 0.755. The slopes were 1.021, 0.939, and 0.977. And the 3- and 5-year AUCs were 0.797 and 0.793, 0.786 and 0.788, and 0.757 and 0.757, respectively.
We were the first to develop and validate exemplary survival prediction models for VC patients and generate corresponding nomograms that allow for individualized survival prediction and could assist clinicians in performing risk-adapted follow-up and treatment.