Personalizing survival predictions in advanced colorectal cancer: The ARCAD nomogram project

Abstract

Background: Estimating prognosis on the basis of clinicopathologic factors can inform clinical practice and improve risk stratification for clinical trials. We constructed prognostic nomograms for one-year overall survival and six-month progression-free survival in metastatic colorectal carcinoma by using the ARCAD database. Methods: Data from 22 674 patients in 26 randomized phase III clinical trials since 1997 were used to construct and validate Cox models, stratified by treatment arm within each study. Candidate variables included baseline age, sex, body mass index, performance status, colon vs rectal cancer, prior chemotherapy, number and location of metastatic sites, tumor mutation status (BRAF, KRAS), bilirubin, albumin, white blood cell count, hemoglobin, platelets, absolute neutrophil count, and derived neutrophil-to-lymphocyte ratio. Missing data (<11%) were imputed, continuous variables modeled with splines, and clinically relevant pairwise interactions tested if P values were less than .001. Final models were internally validated via bootstrapping to obtain optimism-corrected calibration and discrimination C-indices, and externally validated on a 10% holdout sample from each trial (n ¼ 2257). Results: In final models, all included variables were associated with overall survival except for lung metastases, and all but total white cell count associated with progression-free survival. No clinically relevant pairwise interactions were identified. Final nomogram calibration was good (C ¼ 0.68 for overall and C ¼ 0.62 for progression-free survival), as was external validity (concordance between predicted >50% vs < 50% probability) and actual (yes/no) survival (72.8% and 68.2% concordance, respectively, for one-year overall and six-month progression-free survival, between predicted [>50% vs < 50% probability] and actual [yes/no] overall and progression-free survival). Median survival predictions fell within the actual 95% Kaplan-Meier confidence intervals. Conclusions: The nomograms are well calibrated and internally and externally valid. They have the potential to aid prognostication and patient-physician communication and balance risk in colorectal cancer trials.