Machine Learning Models Cannot Replace Screening Colonoscopy for the Prediction of Advanced Colorectal Adenoma

Area of Science:

• Oncology
• Biostatistics
• Medical Informatics

Background:

• Current colorectal cancer (CRC) screening relies on colonoscopy and fecal occult blood testing.
• Non-invasive risk-stratification systems are not yet standard in European guidelines.
• Predicting advanced adenomas (AAs) is crucial for effective CRC screening.

Purpose of the Study:

• To evaluate machine learning (ML) methods for predicting advanced adenomas (AAs) in a CRC screening population.
• To assess the accuracy of logistic regression (LR) and extreme gradient boosting (XGBoost) algorithms.
• To determine if ML models improve risk stratification beyond established factors.

Main Methods:

• Utilized data from 5862 individuals in a CRC screening program.
• Included 36 laboratory parameters, 8 clinical parameters, and 8 dietary patterns.
• Trained and evaluated LR and XGBoost models for AA prediction.

Main Results:

• Moderate accuracy (AUC-ROC 0.65-0.68) was achieved using ML models.
• Including established risk factors did not significantly enhance prediction performance.
• Subgroup analyses (age, genetics, gender) yielded similar results.

Conclusions:

• ML models based on current laboratory and clinical data do not accurately predict advanced adenomas.
• Further research is needed to develop effective non-invasive risk-stratification tools for CRC screening.