Machine Learning-Based Prediction of Histopathological Classification in Colorectal Polyps

Area of Science:

• Gastroenterology
• Oncology
• Data Science

Background:

• Colorectal polyps are precursors to colorectal cancer, necessitating accurate histopathological classification for risk assessment.
• Predicting polyp types aids in early clinical management and personalized screening strategies.
• Machine learning (ML) models offer potential for predicting polyp histopathology using accessible patient data.

Purpose of the Study:

• To evaluate the efficacy of ML algorithms in predicting colorectal polyp histopathological types.
• To identify key demographic, clinical, and dietary predictors of polyp histopathology.
• To explore the potential of ML for individualized colorectal cancer screening.

Main Methods:

• Retrospective analysis of 491 patients undergoing first-time colonoscopy.
• Application of four ML algorithms: decision tree, random forest, support vector machines (SVMs), and extreme gradient boosting.
• Evaluation of model performance using accuracy, sensitivity, specificity, and SHapley Additive exPlanations for variable importance.

Main Results:

• ML models achieved prediction accuracies ranging from 70.9% to 76.4%, with SVM and random forest showing the highest performance.
• The 'no polyp' group was predicted with high accuracy (85.6%-95.9% sensitivity).
• Frequent bulgur consumption, red meat intake, age, and BMI were identified as significant predictors.

Conclusions:

• ML algorithms can accurately predict colorectal polyp histopathological types from routine data.
• This approach supports personalized screening, moving beyond traditional age-based guidelines.
• Integrating ML into screening protocols may enhance early detection and management of colorectal cancer precursors.