Development and validation of a model to predict survival in colorectal cancer using a gradient-boosted machine

Area of Science:

• Oncology
• Medical Informatics
• Biostatistics

Background:

• Accurate prediction of therapy benefit is crucial for successful treatment planning in colorectal cancer (CRC).
• Existing nomograms for CRC often rely solely on tumor-specific features.
• There is a need for models that integrate patient demographics and medical information for comprehensive risk prediction.

Purpose of the Study:

• To develop an accurate and explainable prediction model for 10-year mortality risk following colorectal cancer (CRC) diagnosis.
• To incorporate both tumor characteristics and patient medical/demographic data into the predictive model.
• To enhance treatment planning by providing individualized risk assessments for CRC patients.

Main Methods:

• Utilized data from the Prostate, Lung, Colorectal and Ovarian Cancer Screening (PLCO) Trial, including 154,900 participants.
• Trained a gradient-boosted model on CRC patient data to predict 10-year mortality risk.
• Employed Shapley values to identify the 20 most relevant predictive features and provide model explainability.

Main Results:

• The model was trained and validated on a dataset of 2359 CRC patients, with 686 (29%) deaths from CRC during follow-up.
• Achieved high predictive performance with an area under the receiver operating characteristic curve of 0.84 (±0.04) and accuracy of 0.83 (±0.04).
• The developed model is accessible online for research purposes, facilitating further investigation and application.

Conclusions:

• A robust predictive model for colorectal cancer (CRC) mortality has been developed and validated using prospective data from a large, multicenter cohort.
• The model demonstrates high predictive performance at an individual patient level.
• This tool can significantly aid in clinical decision-making and the discussion of personalized treatment strategies for CRC.