Population-based colorectal cancer risk prediction using a SHAP-enhanced LightGBM model

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Colorectal cancer (CRC) is a leading cause of cancer-related mortality worldwide.
• Early detection and accurate risk stratification are crucial for improving CRC patient outcomes.
• Existing diagnostic and prognostic tools require enhancement for greater precision.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for improved CRC identification.
• To enhance prognostic evaluation using clinical data.
• To create a clinically applicable tool for real-time risk assessment.

Main Methods:

• Utilized multicenter datasets for training, internal, and external validation of ML models.
• Compared seven ML algorithms, selecting Light Gradient Boosting Machine (LightGBM) as the optimal model.
• Employed rigorous performance assessments including AUROC, calibration curves, Brier scores, and SHAP for feature interpretation.

Main Results:

• The LightGBM model achieved high predictive accuracy with AUROCs of 0.9931 (training) and 0.9429 (external validation).
• SHAP analysis identified key predictors, notably age and CA19-9, alongside hematological and biochemical markers.
• A web-based risk calculator was developed for practical clinical use.

Conclusions:

• The developed LightGBM model offers high predictive accuracy and clinical interpretability for CRC.
• The identified biomarker panel provides insights into CRC pathogenesis.
• This tool has significant potential to optimize early CRC diagnosis and personalized risk management.