Machine learning–informed navigation of patients in persistent poverty zip codes to improve colorectal cancer screening: A prospective controlled study.

Abstract

10539 Background: Colorectal cancer (CRC) screening remains suboptimal, particularly in racially diverse and low-income populations. Machine learning (ML) algorithms trained on routine electronic health record (EHR) data may accurately predict CRC risk. Further, patient navigation programs can address screening barriers such as lack of transportation access and low health literacy around colonoscopy preparation. We tested the impact of an intervention combining ML-based risk stratification and an evidence-based CRC screening navigation program on CRC screening rates and outcomes. Methods: This prospective nonrandomized controlled study (NCT05383976) was conducted at a large academic health system, enrolling adults over 50 years of age and at average-risk of CRC. All adults resided in Persistent Poverty zip codes (defined by the US Census Bureau) in the Philadelphia metro area, were attributed to a primary care physician (PCP), and had a colonoscopy order that had not been completed within the past 6 months. We adapted a validated ML algorithm that was previously trained on 22 variables – including age, sex, and longitudinal complete blood counts – to predict CRC risk. Patients enrolled in the intervention arm were prioritized by the ML algorithm for a structured navigation program, including risk-targeted phone-based education, appointment facilitation, transportation support, and mailed FIT tests. A concurrent control cohort received navigation but was not prioritized by the ML algorithm. Adjusted logistic regression models assessed the intervention’s impact on the co-primary outcomes of (1) CRC screening completion (colonoscopy or FIT), and (2) positive screening result, defined as precancerous adenoma (sessile serrated, inflammatory, villous, malignant) and/or positive FIT. Results: 382 patients were enrolled (199 intervention, 183 control). Among intervention patients, 71.4% were reached via phone and 66.4% scheduled screening, with 26.2% and 17.6% completing colonoscopy and FIT, respectively (see Table). Screening completion was similar for intervention vs. control (46.2% vs. 43.2%; adjusted OR 1.02, 95% CI 0.67-1.56, p=0.93). For intervention vs. control, precancerous adenoma detection was 8.5% vs. 5.2% (aOR 2.43, 95% CI 0.48–12.3, p=0.57) and tubular adenoma detection was 35.6% vs. 25.0%. Conclusions: ML-informed navigation was feasible, did not increase screening engagement, and marginally increased rates of precancerous and tubular adenoma detection. Refinements in ML risk stratification and enhanced navigator outreach may maximize impact. Clinical trial information: NCT05383976 . Navigator services provided in intervention group. Number of telephone calls to outreach 1 81 (40.7%) 2 34 (17.1%) 3+ 27 (13.5%) Prep letter sent 11 (7.7%) Prep kit sent 24 (16.9%) Colonoscopy/FIT education 27 (19.0%) Transportation arrangement 8 (5.6%)