Predicting chemotherapy-induced myelosuppression in colorectal cancer: An interpretable, machine learning-based...

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning

Background:

• Colorectal cancer (CRC) treatment often involves chemotherapy, a cornerstone of care.
• Chemotherapy-induced myelosuppression is a significant dose-limiting toxicity in CRC patients.
• Existing predictive models lack interpretability and machine learning integration for myelosuppression risk.

Purpose of the Study:

• To develop and validate an interpretable clinic-machine learning nomogram for predicting chemotherapy-induced myelosuppression in CRC.
• To integrate clinical predictors with machine learning algorithms for accurate risk estimation.
• To provide a tool for individualized prevention strategies in CRC patients undergoing chemotherapy.

Main Methods:

• Retrospective analysis of 855 CRC patients receiving first-line chemotherapy.
• Feature selection using LASSO, decision tree, random forest, and expert consensus identified 10 predictors.
• Ten machine learning algorithms were evaluated; the optimal model was integrated into a nomogram via a feature mapping algorithm.

Main Results:

• The clinic-machine learning nomogram demonstrated high predictive performance (AUC=0.96, AUPRC=0.93) and accuracy (0.90).
• Extreme gradient boosting showed the best initial performance (AUC=0.97).
• The nomogram exhibited good calibration, clinical utility, and robustness in internal testing (AUC=0.95).

Conclusions:

• The developed clinic-machine learning nomogram is a reliable tool for predicting chemotherapy-induced myelosuppression in CRC.
• The nomogram offers clinical interpretability and utility for personalized risk assessment.
• This tool supports optimized treatment decision-making and individualized patient care strategies.