Background: The ARCAD CRC nomogram is designed to predict 1-yr survival in stage 4 CRC to assist with prognostication (Sjoquist, JNCI, 2018). However, this nomogram was developed using patients (pts) enrolled in clinical trials. We sought to externally validate these data in a real-world cohort. Methods: We used the Flatiron database from nationwide records collected from 2013-2020 to create a retrospective Flatiron real-world cohort (FRWC) of pts with histologically confirmed stage 4 CRC. Pts with baseline blood tests prior to treatment (Tx) who had received at least 1 line of Tx were included. Missing data (<8.5%) were imputed using R package MissForest. A multivariable Cox proportional hazards model for overall survival (OS) was applied to fit the reduced model derived from the published ARCAD nomogram as variables from the full model such as number and sites of metastasis were unavailable in Flatiron. The C-index was used to validate the fitted models. The area under a time-dependent ROC (AUROC) analysis was applied to indicate the model prediction performance. Results: A total of 9710 pts, 5740 of whom were deceased were analyzed. The FRWC was older (64 vs 61) and had a poorer ECOG-PS (2+ 14.7% vs 4.2%) compared to the ARCAD. A reduced ARCAD nomogram was derived from the full model (Table). The reduced nomogram calibration was as good as the full model (C-index 0.67 vs 0.68). The 1-yr OS prediction in FRWC using the reduced ARCAD nomogram was good with an AUROC of 0.74. The predictive and observed 1-yr OS using the reduced ARCAD nomogram in the FRWC was 0.640 (95% CI 0.637, 0.644) and 0.727 (95% CI 0.718,0.736). Conclusions: The ARCAD nomogram was validated in the very large FRWC. The observed underestimation of survival by the ARCAD nomogram in the FRWC is likely due to recent advances in Tx options for CRC including targeted- or immunotherapy as the FRWC is more recent compared to pts enrolled in the ARCAD cohort (1997-2012). Our findings indicate that the ARCAD nomogram is a promising decision-making tool for clinicians to predict 1-yr OS in real-world populations.

^†Single HR not available due to nonlinear effect for continuous variables.