Accurate population-based model for individual prediction of colon cancer recurrence

Abstract

Background Prediction models are useful tools in the clinical management of colon cancer patients, particularly when estimating the recurrence rate and, thus, the need for adjuvant treatment. However, the most used models (MSKCC, ACCENT) are based on several decades-old patient series from clinical trials, likely overestimating the current risk of recurrence, especially in low-risk groups, as outcomes have improved over time. The aim was to develop and validate an updated model for the prediction of recurrence within 5 years after surgery using routinely collected clinicopathologic variables.

Material and methods A population-based cohort from the Swedish Colorectal Cancer Registry of 16,134 stage I–III colon cancer cases was used. A multivariable model was constructed using Cox proportional hazards regression. Three-quarters of the cases were used for model development and one quarter for internal validation. External validation was performed using 12,769 stage II–III patients from the Norwegian Colorectal Cancer Registry. The model was compared to previous nomograms.

Results The nomogram consisted of eight variables: sex, sidedness, pT-substages, number of positive and found lymph nodes, emergency surgery, lymphovascular and perineural invasion. The area under the curve (AUC) was 0.78 in the model, 0.76 in internal validation, and 0.70 in external validation. The model calibrated well, especially in low-risk patients, and performed better than existing nomograms in the Swedish registry data. The new nomogram’s AUC was equal to that of the MSKCC but the calibration was better.

Conclusion The nomogram based on recently operated patients from a population registry predicts recurrence risk more accurately than previous nomograms. It performs best in the low-risk groups where the risk-benefit ratio of adjuvant treatment is debatable and the need for an accurate prediction model is the largest.