A predictive model of longitudinal, patient-specific colonoscopy results

Eric A. Sherer, Sanmit Ambedkar, Sally Perng, Yuehwern Yih, Thomas F. Imperiale

Research output: Contribution to journalArticle

5 Scopus citations

Abstract

We suggest a model framework, in which an individual patient's risk for colonic neoplasia varies based on findings from his previous colonoscopies, to predict longitudinal colonoscopy results. The neoplasia natural history model describes progression through four neoplasia development states with patient age. Multiple natural history model parameter sets are assumed to act concurrently on the colon and parameter set prevalence combinations, whose a priori likelihoods are a function of patient sex, provide a basis set for patient-level predictions. The novelty in this approach is that after a colonoscopy, both the parameter set combination likelihoods and their model predictions can adjust in a Bayesian manner based on the results and conditions of the colonoscopy. The adjustment of model predictions operationalizes the clinical knowledge that multiple or advanced neoplasia at baseline colonoscopy is an independent predictor of multiple or advanced neoplasia at follow-up colonoscopy - and vice versa for negative colonoscopies - and the adjustment of parameter set combination likelihoods accounts for the possibility that patients may have different neoplasia development rates. A model that accurately captures serial colonoscopy results could potentially be used to design and evaluate post-colonoscopy treatment strategies based on the risk of individual patients. To support model identification, observational longitudinal colonoscopy results, procedure details, and patient characteristics were collected for 4084 patients. We found that at least two parameter sets specific to each sex with model adjustments was required to capture the longitudinal colonoscopy data and inclusion of multiple possible parameter set combinations, which account for random variations within the population, was necessary to accurately predict the second-time colonoscopy findings for patients with a history of advanced adenomas. Application of this model to predict CRC risks for patients adhering to guideline recommended follow-up colonoscopy intervals found that there are significant differences in risk with patient age, gender, and preparation quality and demonstrates the need for a more rigorous investigation into these recommendations.

Original languageEnglish (US)
Pages (from-to)563-579
Number of pages17
JournalComputer Methods and Programs in Biomedicine
Volume112
Issue number3
DOIs
StatePublished - Dec 1 2013

    Fingerprint

Keywords

  • Biomedical engineering
  • Colorectal cancer
  • Dynamic simulation
  • Mathematical modeling
  • Patient specific
  • Population balance

ASJC Scopus subject areas

  • Computer Science Applications
  • Software
  • Health Informatics

Cite this