EP451 - Evaluation Of Machine Learning Methods To Reduce Diagnostic Delay In Young Adult Rectal And Anal Cancer Patients

Purpose/Background: Young adults are at risk of delayed diagnosis of rectal and anal cancer, due to fragmented healthcare access with poor documentation, prior authorization requirements for screenings, and lower clinical index of suspicion. Machine learning models that predict cancer risk based on demographic and lifestyle factors exist, but either provide baseline risk or require extensive institution-specific data. Widely applicable models that provide continuously updated, patient-specific risk assessments, even in patients receiving fragmented or suboptimal clinical care, are greatly needed. This study aims to evaluate a machine learning model that uses ICD code sequences to triage patients who are at risk of future rectal or anal cancer diagnosis.

Methods/Interventions: This is a retrospective case-control study. 128 rectal/anal cancer patients and 896 age and gender matched patients with non-malignant anorectal diseases from a single urban and rural hospital network were included. All patients were diagnosed after January 1, 2015 and were aged 18-49 at the time of diagnosis. Up to 3 years of ICD-9 and -10 codes preceding initial diagnosis were obtained; the frequency and timing of code occurrences were vectorized for each patient. An XGBoost machine learning model was trained to predict whether a patient would be diagnosed with rectal or anal cancer in the next 90 days, based only on their vectorized ICD codes. The model was trained and evaluated using 10-fold cross-validation. Feature importance was calculated using SHAP values.

Results/Outcomes: 128 patients with rectal or anal cancer (median age 45.0y, 53.9% female, 88% Caucasian) and 896 patients with non-malignant anorectal disease (median age 44.9y, 53.9% female, 89% Caucasian) were studied. The model predicted cancer diagnosis 90 days prior to physician-documented diagnosis with a specificity of 99.6% (standard deviation: 0.6%), recall/sensitivity 43.8% (13.9%), precision 94.4% (7.3%), and F1 score 0.58 (0.15). ICD codes significantly predicting impending anorectal cancer diagnosis included a 6-month history of benign anorectal conditions (ICD-10 chapter K62), colon cancer (C18), or various gastrointestinal conditions like hematemesis and melena (K92), among others.

Conclusion/Discussion: A machine learning model distinguishes young adults with rectal and anal cancer from ultimately nonmalignant anorectal disease with a lead time of 90 days. Using this model, approximately one in 2.4 rectal/anal cancer patients would be diagnosed at least 90 days earlier. The model only requires billed ICD code history: it does not require medical records or patient questionnaires, demonstrating the feasibility of using this highly portable and easily accessible data type to reduce triage barriers in common colorectal presentations. The high specificity of this model and overall performance in a small dataset are promising for validation in large populations, and future use in patient triage to appropriate colorectal care.