V70: Application of Machine Learning to Determine United States County-level Health Predictors of Disparities in Gastrointestinal Cancer Mortality Trends

Monday, February 28, 2022

6:16 PM – 6:19 PM CST

Location: Zoom Meeting

Virtual Poster Presenter(s)

Taylor J. Jaraczewski, MS, MD

General Surgery Resident
Medical College of Wisconsin
Milwaukee, Wisconsin, United States

Disclosure: Disclosure information not submitted.

Participants should be aware of the following financial/non-financial relationships:

Taylor J. Jaraczewski, MS, MD: Disclosure information not submitted.

Introduction: Rates of death from gastrointestinal (GI) cancers have decreased across the United States (U.S.) over time; however, disparities in the magnitude of improvement across all geographic regions of the country are not well characterized. The objective of this study was to identify U.S. counties with rising cancer mortality and use machine learning to create a phenotype of county-level health factors that define these counties.

Methods: County-level cancer outcomes data were obtained from the National Cancer Institute State Cancer Profiles database and linked to data from the County Health Rankings. Cancer-specific mortality (5-year trend) was obtained for all U.S. counties for colon and rectum, esophageal, liver and bile duct, and stomach cancers between 2015 - 2019. Excluded were counties with incomplete health rankings or mortality data. Counties with at least 1 or more cancers with a trend of worsening 5-year mortality were designated as “high risk”. Partial least squares regression (PLSR) was utilized to predict high risk counties using county level health factors. All analysis were conducted using Matlab and R.

Results: Overall there was a 1.3% decrease in GI cancer mortality in the U.S. A total of 81/1886 (4.3%) counties were designated as “high risk” with an average rise in mortality of 1.7%. The best-fit PLSR model utilized 10 principal components and explained 40.1% of the variance between high and normal risk counties. Variable importance of project scores identified the following county-level health factors as most influential for predicting high risk counties: percentage of population with access to places for physical activity, number of deaths, alcohol-impaired driving deaths, driving deaths, dentists, unemployed, uninsured, single parent households, deaths from injuries, and the number of public associations.

Conclusions:

Despite overall decreasing GI cancer-related mortality in the U.S, some counties have rising rates. These high-risk counties can be identified using network phenotypes that include insurance, infrastructural, social and educational county-level factors providing targets for intervention and improvement.

Learning Objectives:

describe the utilization of machine learning for analyzing multifactorial problems.
list county health factors associated with worsening 5 year gastrointestinal cancer trends.
demonstrate an understanding that though cancer outcomes overall are improving in the United States, sub populations exist with stagnant or worsening trends.