Curie Brief
Turn on cookies to sign in
Signing in saves your progress to your Curie account. We can only do that with cookies on — turn them on to continue.
News
Two AI models agree 97.5% of the time on spotting colorectal neoplasia across 5,000 colonoscopies, but there's a catch. The more sensitive random forest model came with a much higher false-positive rate compared to a large language model. Researchers say more prospective studies are needed before either tool is ready for routine clinical use.
Researchers compared two machine learning models—a random forest model and a large language model (LLM)—to see how well they could automate adenoma detection rate (ADR) reporting from colonoscopy records. Analyzing 5,000 colonoscopy reports from 2023 to 2025, the study found that both models agreed on colorectal neoplasia classification in 97.5% of cases, a strong signal that AI could streamline a critical quality metric in colorectal cancer prevention.
But the devil's in the details. Among the 2.5% of cases where the models disagreed, the random forest model was more sensitive—catching more adenomas—but also flagged far more false positives. The LLM, meanwhile, was more precise but came with higher implementation costs and missed slightly more lesions in certain categories.
By the Numbers:
Why it matters: Accurate ADR reporting is a key benchmark for colonoscopy quality and colorectal cancer prevention. AI-assisted automation could reduce manual burden and standardize reporting—but clinicians need to weigh sensitivity against false-positive risk and cost before adopting either model at scale.
