A Machine-Learning Approach to Predicting Adenoma Detection and Surveillance Impact of Deeper Sections in Colorectal Polypectomy Specimens.

Abstract

In the routine practice of pathology, cutting deeper levels into paraffin blocks of colorectal polypectomy specimens that do not show a lesion on initial hematoxylin-eosin sections is common. However, this practice lacks standardization, and data are limited on the predictability of adenoma detection and the impact on surveillance intervals and resource use. To develop machine learning models that predict which polyps and encounters are most likely to reveal clinically significant lesions on deeper levels and influence patient surveillance intervals. We performed a retrospective study of 94 888 patients (145 405 colonoscopies; 392 130 polyps) from 2007 to 2024. Polyp characteristics, deeper-level requests, and clinical/endoscopic features were analyzed. Machine learning models were developed to predict (1) adenoma detection in single polyps with deeper levels and (2) impact on surveillance intervals per encounter. Deeper levels were requested in 26 067 jars (11%) and 21 232 encounters (15%), revealing adenomas in 7946 single polyps examined (51%). Predictive models showed high performance: logistic regression for single-polyp conversion (area under the receiver operating characteristic curve 0.88, accuracy 81%) and gradient boosting classifier for surveillance interval impact (area under the receiver operating characteristic curve 0.90, accuracy 83%). Important predictors included polyp size, location, total polyps, and artificial intelligence-assisted colonoscopy. Deeper-level sectioning contributes significantly to adenoma detection and influences surveillance intervals. Machine learning provides a practical framework to optimize deeper-level requests, enhancing diagnostic precision and resource allocation.