Using Machine Learning to Leverage Biomarker Change and Predict Colorectal Cancer Recurrence
Patricia J Rodriguez1, Patrick J Heagerty2, Samantha Clark1
1The Comparative Health Outcomes, Policy & Economics (CHOICE) Institute, University of Washington, Seattle, WA.
JCO Clinical Cancer Informatics
|November 14, 2023
Summary
A machine learning model using carcinoembryonic antigen (CEA) levels effectively predicts colorectal cancer (CRC) recurrence. Tracking changes in CEA over six months offers a simple yet powerful tool for personalized risk assessment in CRC patients.
Area of Science:
- Oncology
- Biomarkers
- Machine Learning
Background:
- Colorectal cancer (CRC) recurrence risk varies significantly among patients and over time.
- Personalized risk prediction can improve patient management and decision-making.
- Utilizing longitudinal patient data, including biomarker profiles, is key to refining risk assessments.
Purpose of the Study:
- To develop and evaluate a predictive model for colorectal cancer recurrence.
- To investigate the utility of longitudinal carcinoembryonic antigen (CEA) biomarker data in risk prediction.
- To compare the performance of various machine learning algorithms for predicting CRC recurrence.
Main Methods:
- Utilized electronic health record data from 3,970 patients diagnosed with stage I-III CRC.
- Engineered features from longitudinal CEA measurements to capture temporal trends.
- Employed a discrete-time Superlearner model to assess multiple algorithms for recurrence prediction.
Main Results:
- The XGBoost with depth = 1 (XGB-D1) model demonstrated superior performance in predicting recurrence (AUC 0.87-0.94).
- The 6-month change in log CEA was the sole predictor variable identified by the optimal XGB-D1 model.
- Risk stratification based on CEA change showed distinct recurrence probabilities, highlighting the model's predictive power.
Conclusions:
- A machine learning model incorporating longitudinal CEA data provides a simple and highly effective method for predicting CRC recurrence.
- The 6-month change in log CEA is a critical and sufficient feature for accurate recurrence prediction.
- This approach enhances personalized risk assessment and decision-making for colorectal cancer patients.
