(373e) Colorectal Cancer Risks at Follow-up Colonoscopy: Incorporating Initial Colonoscopy Findings Into a Longitudinal Adaptive-Predictive Model

Mathematical models of colorectal cancer (CRC) development have been used to design screening and surveillance strategies using procedures such as colonoscopy (Loeve et al. 1999; Roberts et al. 2007), but these models only consider the risk factors for neoplasia at the initial colonoscopy. Screening models often stratify the population according to patient demographic and behavior, but current clinical colonoscopy surveillance guidelines are based on risk factors such as: multiple adenomas, large adenomas, and advanced histology ? where the clinical variables' adjusted odds ratios for developing subsequent advanced neoplasia can be significantly larger than those of demographic and behavioral variables (Martinez et al. 2009). We demonstrate that accounting for the known surveillance clinical risk factors, by using information contained in an initial colonoscopy, improves the accuracy of mathematical models for predicting neoplastic findings at follow-up colonoscopies.

The mathematical frame consists of a series of CRC development trajectories and the likelihoods that an individual patient follows a given trajectory. Initially these likelihoods are based on demographics and behaviors (so the screening predictions match those of previously developed models) but adjust in a Bayesian manner according to the colonoscopic indications and findings. To test the model, we collected longitudinal colonoscopy information (i.e. number of adenomas, sizes, and locations; preparation quality; and age at procedure) from over 1,500 patients who had their first colonoscopy performed at the Roudebush VA Medical Center.

We found that the adaptive model improves the overall predictions for the number of adenomas at surveillance colonoscopies (The R2 value decreases by over 20% and average maximum likelihood improvement was five times larger) as well as for the individual (the predictions for over two-thirds of the patients are more accurate with the adaptive model than for demographic stratification only and the median maximum likelihood estimate was 1.22 times larger). In particular, the increase in accuracy of the adaptive model is focused in two patients groups: patients with numerous adenomas and those with none or very few adenomas.

In summary, this work utilizes a mathematical framework that varies the model parameters according to the relevant clinical findings at colonoscopy in a Bayesian manner so the subsequent model predictions are adjusted based on these findings. These adjustments do show an increase in accuracy for follow-up predictions relative to models based solely on patient demographic and behavioral stratification.

REFERENCES:

Loeve F, Boer R, van Oortmarssen GJ, van Bellegooijen M, and Habbema JDF. ?The MISCAN-COLON Simulation Model for Evaluation of Colorectal Cancer Screening?, Computers and Biomedical Research, 32: 13-33, 1999.

Martinez ME, Baron JA, Lieberman DA, Schatzkin A, Lanza E, Winawer SJ, Zauber AG, Jiang R, Ahnen DJ, Bond JH, Church TR, Robertson DJ, Smith-Warner SA, Jacobs ET, Alberts DS, and Greenberg ER. ?A Pooled Analysis of Advanced Colorectal Neoplasia Diagnoses After Colonoscopic Polypectomy?, Gastroenterology, in press ? available online, 2009.

Roberts S, Wang L, Klein R, Ness R, and Dittus R. ?Development of a simulation model of colorectal cancer?, Transactions on Modeling and Computer Simulation, 18: 1-30, 2007.