Related Experiment Video
Updated: Jun 25, 2026
Comparison of Predictive Performance of Three Lymph Node Staging Systems in Colorectal Signet Ring Cell Carcinoma Based on Machine Learning Model
Published on: April 18, 2025
Predicting High-Risk Colorectal Polyps in African Americans Using Pre-colonoscopy Clinical Features: Machine Learning
Basheer Qolomany1, Mrinalini Deverapall2, Adeyinka Laiyemo2
1Departments of Internal Medicine, Pathology, and Biochemistry, and Cancer Center, Howard University College of Medicine, Washington, D.C., 20059, USA. Basheer.Qolomany@Howard.edu.
Predicting high-risk colorectal polyps using noninvasive, pre-colonoscopy data is feasible but shows limited generalizability. Machine learning models identified key demographic and clinical factors, highlighting potential for risk stratification, especially in diverse populations.
Area of Science:
- Colorectal cancer research
- Machine learning in healthcare
- Epidemiology of gastrointestinal diseases
Background:
- Traditional risk stratification for advanced colorectal polyps relies on colonoscopy and pathology.
- Noninvasive, pre-colonoscopy features offer potential for enhanced clinical decision-making and equitable risk assessment.
- Identifying high-risk patients pre-colonoscopy can optimize resource allocation and reduce unnecessary procedures.
Purpose of the Study:
- To develop and externally validate machine learning models for predicting high-risk colorectal polyps (HRP).
- To utilize only noninvasive, pre-colonoscopy demographic, clinical, and behavioral features.
- To assess model performance in a diverse, urban cohort, predominantly African American.
Main Methods:
- A retrospective cohort study involving 4,681 patients for internal validation and 1,562 for external validation.
- Development and comparison of multiple machine learning models (neural networks, random forest, SVM, Naïve Bayes, logistic regression, decision trees, KNN, XGBoost).
- High-risk polyps defined by histology (villous/tubullovillous adenomas, high-grade dysplasia), size (≥10 mm), or number (≥3). Performance evaluated using ROC-AUC, PR-AUC, accuracy, precision, recall, and F1 score; interpretability via SHAP.
Main Results:
- Overall predictive performance using noninvasive features was moderate.
- Neural networks showed highest internal performance (ROC-AUC 0.78) but poorer external validation (ROC-AUC 0.67).
- Simpler models (Naïve Bayes, SVM, XGBoost) had lower internal performance but more stable external generalization (ROC-AUC ~0.52-0.63). Key predictors included age, smoking status, sex, occupation, race, indication for colonoscopy, and family history.
Conclusions:
- Prediction of high-risk colorectal polyps using routine pre-colonoscopy data is feasible but exhibits limited generalizability.
- Findings underscore the clinical potential and limitations of pre-procedural risk modeling, particularly in diverse, underserved populations.
- Integration of additional data modalities may be necessary for robust and equitable prediction tools.
