Page 38 - Risk quantification and modification in older patients with colorectal cancer
P. 38
Chapter 2
Technical analysis and performance
Tables 4A-C shows the technical analysis and performance of all model. For mortality (Table 4A), the reported discrimination (AUC) of the models during model development ranged from 0.73 to 0.91. However, eight studies did not report confidence intervals.11-14,16,18,23,26 The internal validation methods that were reported were random-split,10,11,18,26 cross-validation,13 two studies reported external validation.16,20
Additional external validation studies were found for the following models: CR-POSSUM,39-42 CCF-CRM,43 AFC Index,44 CR-BHOM,41 revised ACPGBI,45 and IRCS.42 Discrimination ranged from 0.56-0.89. However, calibration was considered poor except for the CR-BHOM and AFC model. Calibration could not be judged for the ACS-NSQIP model13. The performance of the CR-POSSUM10 and CR-BHOM14 models in a cohort of 991 Portuguese octogenarians were AUC 0.74 and 0.65, and poor and good calibration, respectively.41
For the anastomotic leakage models (Table 4B), discriminaton of the models ranged from 0.63 to 0.95 (in the development cohort). Discrimination in the development phase did not apply to the studies of Dekker et al.15 and Rojas- Machado et al.27 because of their meta-analysis approach for model development. Remarkably, the study of Pasic et al. reported an AUC of 1.0 (validation) without a confidence interval in a small study population of 40 patients.19 Additional external validation studies were found for the CLS,27,46 anastomoticleak.com,46 and ACS-NSQIP anastomotic leakage model.46 In these studies, AUC ranged from 0.58-0.80. Calibration could not be judged for the ACS-NSQIP anastomotic leakage model and JSCCR model.
For the “other” outcomes, the methodological and model performance analysis is shown in Table 4C. Of note, the discriminatory performance of the COLA-score model varied across the countries France, England and Switzerland (AUC 0.60- 0.64), with poor calibration in all three cohorts.
Online calculators can be found for the CR-POSSUM10, ACS-SNQIP13,26,29 and ACPGBI models.16 Other modes of presentation were a formula, chart or nomogram.
36
