QS218 - Predicting Early-Onset Colorectal Cancer in Individuals Below Screening Age Using Machine Learning and Real-World Data

Quick Shot Presenter(s)

• 

  Isabela Visintin, MD

  General Surgery Resident
  UF Health Jacksonville
  Jacksonville, FL, United States

  All of the relevant financial relationships listed below have been mitigated.
  This individual has no financial relationships with ineligible companies.

Author(s)

• CS

  Chengkun Sun, n/a

  student
  University of Florida
  Gainesville, Florida, United States

  Disclosure information not submitted.
• EM

  Erin M. Mobley, PhD, MPH

  Assistant Professor
  University of Florida
  Jacksonville, Florida, United States

  Disclosure information not submitted.
• ZA

  Ziad T. Awad, M.D, FACS

  Chief of Gastrointestinal Surgery and Surgical Oncology
  University of Florida Health at Jacksonville
  Jacksonville, Florida, United States

  Disclosure information not submitted.
• JF

  Jennifer Fishe, MD

  Associate Professor
  University of Florida College of Medicine - Jacksonville
  Jacksonville, Florida, United States

  Disclosure information not submitted.
• AP

  Alexander Parker, n/a

  Senior Associate Dean
  University of Florida
  Jacksonville, Florida, United States

  Disclosure information not submitted.
• TG

  Thomas J. George, MD, FACP, FASCO

  Professor and Deputy Director
  University of Florida Health Cancer Center
  Gainesville, Florida, United States

  Disclosure information not submitted.
• JB

  Jiang Bian, PhD

  Professor and Chief Data Scientist
  University of Florida
  Gainesville, Florida, United States

  Disclosure information not submitted.
• JX

  Jie Xu, n/a

  Assistant Professor
  University of Florida
  Gainesville, Florida, United States

  Disclosure information not submitted.
• MD

  Meghan Daly, MD

  Colon and Rectal Surgeon, Clinical Assistant Professor, Associate Program Director
  University of Florida Health Jacksonville
  Jacksonville, Florida, United States

  Disclosure information not submitted.

Purpose/Background: Colorectal cancer (CRC) is now the leading cause of cancer-related deaths among young Americans. The increasing incidence of early-onset colorectal cancer (EOCRC) in individuals younger than the age of 50 years has led to a lowering of the screening age to 45. Still, patients with EOCRC tend to present at later stages and have lower disease-specific survival rates due to poorly defined risk factors and delays in diagnosis. The rapid integration of artificial intelligence (AI) and big data analytics has significantly expanded the horizons of medical research and clinical care. In the context of EOCRC, the use of machine learning (ML) algorithms for detection of CRC have been employed to predict the risk of the disease, however previous studies faced challenges in effectively matching cases and control groups and failed to distinguish between colon and rectal cancers. As a result, there is a pressing need for improved methods to enhance the reliability and understanding of ML models in EOCRC prediction. Our study aims to predict EOCRC using ML with structured electronic health record (EHR) data for individuals below the screening age of 45 years.

Methods/Interventions: We identified a cohort of patients under 45 with and without an EOCRC diagnosis from the OneFlorida+ Clinical Research Consortium, with encounters from 2012 to 2022. Given the distinct pathology of colon cancer (CC) and rectal cancer (RC), we created separate prediction models for each cancer type with various ML algorithms. Our final study cohort comprised of 1,358 CC cases with 25,485 controls and 560 RC cases with 22,648 controls. We used an incident matching process to match cases and controls to ensure a fair comparison across these groups. We assessed multiple prediction time windows (0, 1, 3, and 5 years) and ensured robustness through propensity score matching (PSM) to account for confounding variables. Model performance was assessed using established metrics. Additionally, we employed Shapley Additive exPlanations (SHAP) to identify risk factors for EOCRC.

Results/Outcomes: Our study yielded Area Under the Curve (AUC) scores of 0.811, 0.748, 0.689, and 0.686 for CC prediction, and 0.829, 0.771, 0.727, and 0.721 for RC prediction at 0, 1, 3, and 5 years, respectively. Notably, predictors of EOCRC included immunodeficiencies such as IBD and HIV, gastrointestinal disorders, secondary cancers and iron-deficiency anemia. Being underweight was a significant symptom associated with both CC and RC. Conversely, obesity, overweight and normal weight were negatively associated with RC development.

Conclusion/Discussion: This study highlights the potential of ML techniques in leveraging EHR data to predict EOCRC, offering valuable insights for potential early diagnosis in patients who are below the recommended screening age. Future investigation should focus on refining the experimental design, expanding data collection to include outpatient population and enhancing performance of the prediction models. Ultimately, these efforts will contribute to early detection and better management of EOCRC, with the goal to improve patient outcomes.

SHAP summary plot for features in colon cancer prediction
Figure S1. SHAP summary plot of the top 20 features in CC prediction using XGBoost models with 1-year and 5-year prediction windows: (A) excluding CRC-related features; (B) excluding cancer-related features. The prefix before the “_” in the y-axis labels of plots indicates the source of the corresponding features in the PCORnet data model. Specifically, these sources are: Diagnosis (Diag), Procedure (Proc), Medication (Med), Vital Signs (Vital), and Demographics (Demo).