Two machine learning approaches demonstrated 97.5% agreement in identifying colorectal neoplasia across 5,000 colonoscopy reports, according to research presented at the 2026 American College of Gastroenterology (ACG) Annual Scientific Meeting. The retrospective study compared a random forest model with a large language model (LLM) for automated reporting of adenoma detection rate (ADR).
Among 123 discordant cases requiring manual review, the random forest model missed two of 16 tubular adenomas, compared with seven missed by the LLM. Both models missed four of 47 sessile serrated lesions. However, the random forest model generated substantially more false positives, incorrectly classifying 92 cases as tubular adenomas compared with eight for the LLM.
The findings highlight a trade-off between sensitivity and false-positive classification. Although the random forest model detected more tubular adenomas in the discordant cases, false positives could artificially inflate ADR, a key measure of colonoscopy quality. The LLM had fewer false positives but potentially higher implementation costs.