Predicting High-Risk Colorectal Polyps in African Americans Using Pre-Colonoscopy Clinical Features: Machine Learning Model Development and Temporal Validation

Technical Summary: Predicting High-Risk Colorectal Polyps in African Americans Using Pre-Colonoscopy Clinical Features

Problem Statement
Colorectal cancer (CRC) remains a leading cause of morbidity and mortality, disproportionately affecting racial and socioeconomic groups. While colonoscopy is the gold standard for prevention, its diagnostic yield varies based on procedural and patient factors, and resources are often limited or unevenly distributed. Traditional risk stratification relies on broad demographic and clinical variables (e.g., age, family history) but often lacks the granularity to identify individual high-risk patients, leading to both overuse and underuse of screening. Furthermore, existing machine learning (ML) models for CRC risk often lack robust validation in diverse, real-world populations, particularly among African Americans, and frequently rely on invasive or post-procedural data. The study addresses the need for non-invasive, pre-colonoscopy risk prediction tools that utilize routinely available clinical data to prioritize surveillance and improve equity in underserved urban populations.

Methodology
The authors conducted a retrospective cohort study at Howard University Hospital involving 6,243 patients who underwent colonoscopy between 2015 and 2024. The dataset was temporally split to ensure rigorous validation:

• Internal Cohort (2015–2022): 4,681 patients used for model development and internal validation.
• External Cohort (2023–2024): 1,562 patients used as a temporally held-out validation set to assess generalizability.

Data and Features:
The study utilized only non-invasive, pre-colonoscopy features extracted from electronic health records (EHR). Intra- and post-procedural variables (e.g., bowel preparation quality, withdrawal time) were explicitly excluded. Features included:

• Demographics: Age, sex, race, occupation.
• Lifestyle: Smoking status, alcohol use, illicit drug use.
• Clinical History: Family history of CRC (1st and 2nd degree), comorbidities (CAD, diabetes, hypertension, IBD, CKD, dialysis, H. pylori, HPV), and medication use (aspirin, NSAIDs, multivitamins, chemo/radiation history).
• Indication: Colonoscopy indication.

Target Variable:
Patients were classified as having High-Risk Polyps (HRP) if they met any of the following criteria: villous/tubulovillous adenomas, high-grade dysplasia, polyp size ≥10 mm, or ≥3 polyps per procedure. All others were classified as Low-Risk Polyps (LRP).

Model Development:
The authors developed and compared eight supervised learning algorithms: Logistic Regression, Random Forest, XGBoost, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Naïve Bayes, Decision Tree, and a Neural Network (NN).

• Preprocessing: Missing values were imputed using the median strategy; continuous variables were standardized. Synthetic Minority Oversampling Technique (SMOTE) was applied within training folds to address class imbalance.
• Training: Internal validation used stratified 5-fold cross-validation. External validation involved retraining models on the full internal cohort and testing on the 2023–2024 set.
• Evaluation Metrics: Performance was assessed using ROC-AUC, PR-AUC, accuracy, precision, recall, and F1 score.
• Interpretability: SHapley Additive exPlanations (SHAP) were used to quantify feature contributions and identify key predictors.

Key Results

• Model Performance: The Neural Network achieved the highest internal discrimination (ROC-AUC 0.78, PR-AUC 0.75, Accuracy 0.72). However, its performance dropped significantly in the external cohort (ROC-AUC 0.67, Accuracy 0.66), suggesting potential overfitting or temporal feature drift.
• Generalization: Simpler models (Naïve Bayes, SVM, XGBoost, Logistic Regression) showed lower internal performance (ROC-AUC 0.54–0.59) but demonstrated more stable generalization to the external cohort (ROC-AUC 0.52–0.63; Accuracy ~0.53–0.60).
• Predictors: SHAP analysis identified age, smoking status, sex, occupation, race, colonoscopy indication, and family history of CRC as the most influential predictors. Traditional clinical factors remained dominant, with sociodemographic variables also contributing meaningfully.
• Cohort Characteristics: The study population was predominantly African American (~84%). While the internal and external cohorts were broadly similar, statistically significant differences were observed in age, sex distribution, smoking prevalence, hypertension, and chronic kidney disease, likely due to the large sample size.

Significance and Claims
The paper claims that predicting high-risk colorectal polyps using routine pre-colonoscopy data is feasible but demonstrates limited generalizability across temporally distinct cohorts.

• Clinical Potential: The study highlights the potential of pre-procedural risk modeling to enhance clinical decision-making, particularly in resource-limited settings, by prioritizing surveillance for high-risk individuals without requiring invasive biomarkers or imaging.
• Equity: By focusing on a predominantly African American, urban cohort—an underrepresented group in prior CRC AI studies—the work addresses a critical gap in understanding model performance in historically underserved populations.
• Limitations of ML: The authors modestly conclude that while ML models can integrate multidimensional data, the incremental benefit over simpler risk scoring strategies may be limited in diverse real-world populations. The decline in performance from internal to external validation underscores the challenges of temporal drift and the need for robust external validation.
• Future Needs: The study suggests that achieving clinically robust and equitable prediction tools will likely require the integration of additional data modalities and prospective validation.

In summary, the paper provides a validated framework for pre-colonoscopy risk stratification using non-invasive data but cautions that current models, while promising, require further refinement and validation before widespread clinical deployment.