3 Machine learning approaches to predict colorectal cancer screening uptake in real-world EHR data

Objectives/Goals: The study’s objective was to use real-world data to compare machine learning approaches for predicting colorectal cancer (CRC) screening status: (1) a Random Forest model optimized for overall accuracy and (2) Feasible Solution Algorithm (FSA) models focused on interpretability and identifying subgroups at risk of being unscreened. Methods/Study Population: The data included 22,549 adults aged 45–75 who visited a UK King’s Daughters Medical Center (UKKD) primary care site between 12/1/2023–11/30/2024, across five Appalachian counties. Patients with total colectomy, colon cancer, advanced illness, hospice care, dementia, and/or frailty were excluded. After complete-case analysis, 18,533 records remained and were split 80/20 for training/testing. Two machine learning methods, random forests and FSA, were applied to identify key predictors of CRC screening status. Within the FSA framework, logistic regression models were fitted using the training data, and model performance metrics were subsequently evaluated using the testing dataset. Results/Anticipated Results: Overall (n=22,549), 65.6% of patients completed CRC screening. The mean age was 59 years (SD = 8.2); most were White (97%), 40% had Medicare insurance, and 68% had multiple non-primary care provider (PCP) visits. The Random Forest model demonstrated the highest overall accuracy of 69% and had notably high sensitivity, correctly identifying approximately 9 of 10 screened patients. On the other hand, the FSA model offered greater specificity (65%) and interpretability, with the 2-predictor model identifying smokers and those with 0 non-PCP visits as the vulnerable subgroups who are least likely to be screened. Discussion/Significance of Impact: FSA identified smoking status, insurance type, number of non-PCP clinic visits, and age as key factors relating to CRC screening behavior. These findings can help clinicians and public health teams identify unscreened patients, guide targeted outreach, and advance translational strategies to improve population health.