Identifying risk factors of late-stage diagnosis of colorectal cancer from unexplained relationships between screening and stage at diagnosis at a small geographic scale

Keywords: regionalization, random forest, machine learning, risk factors, cancer, health disparity
Abstract Type: Poster Abstract

Abstract

High rates of colorectal cancer screening do not always translate into high rates of early-stage diagnosis at the local level. The co-occurrence of high rates of screening and low rates of early-stage diagnosis in some geographic areas may be explained by certain colorectal cancer risk factors in those areas. In this study, we aim to examine the geographic association of colorectal cancer screening and early-stage at diagnosis at a small geographic scale and identify risk factors of late-stage colorectal cancer in areas with unexpected relationships between screening and stage at diagnosis. Individuals newly diagnosed with colorectal cancer from 2010 to 2019 were obtained from the Ohio Cancer Incidence Surveillance System. To prevent re-identification of patients from a small number of cases at the census tract level, we combined adjacent, socio-demographically similar census tracts into the smallest geographic areas that included a minimum number of early or late-stage cases (n=11) using the Max-p-regions method. Next, we used geographically weighted regression (GWR) to identify areas with non-positive and non-significant associations between early-stage diagnosis rates and screening rates. We then examined the variable importance of demographic, environmental, behavioral, socioeconomic, and clinical care risk factors in predicting the proportion of late-stage diagnosis using random forest analysis in those areas. The identification of important community-level risk factors in those areas will help future studies in identifying carcinogens potentially causing aggressive tumors that lead to late-stage diagnosis regardless of high screening uptake.