Large Language Model Outperformed Random Forest in Colonoscopy AI
A new research study found that LLMs may improve diagnostic accuracy for colorectal lesion classification over traditional ML.
Updated on Oct. 13, 2026 in Artificial Intelligence
Live Poll
Do you trust artificial intelligence to accurately report results from medical diagnostic procedures?
Researchers compared a large language model and a random forest model across 5,000 colonoscopy cases to evaluate diagnostic accuracy. The study, presented on October 13, 2026, found that the models reached concordance in 97.5% of instances, though the LLM outperformed its counterpart in cases of disagreement.
Why it matters
Automating diagnostic classification in colonoscopies could significantly improve the scalability of quality monitoring in gastroenterology. Identifying which AI architecture provides the highest accuracy is a critical step before these tools move from research-stage analysis to clinical integration.
Across 5,000 cases, the large language model and random forest model agreed in 4,877 instances, marking a 97.5% concordance rate. In the 123 cases where classifications diverged, the LLM aligned more closely with expert human assessments for both adenomas and serrated lesions.
The players
Mayo Clinic
A medical research and clinical care institution that investigates AI applications to improve diagnostic accuracy and healthcare efficiency.
The details
The study utilized a large language model — a system trained on massive datasets to interpret and generate complex patterns — and a random forest model — an ensemble learning technique that builds multiple decision trees to produce a classification. Researchers assessed the models by comparing their outputs against manual expert reviews of colonoscopy findings. The random forest model displayed a higher rate of false-positive results compared to the LLM, particularly when identifying adenomas and serrated lesions.
Timeline
Colonoscopies analyzed in the study were performed between 2023 and 2025.
Study findings were presented at the ACG Annual Meeting on October 13, 2026.
The Tech Race
This research positions large language models as potential challengers to traditional decision-tree algorithms in medical imaging and classification. It marks a shift from relying on standard random forest models toward generative architectures that may offer higher precision in diagnostic tasks.
This development represents research-stage work and does not currently affect clinical colonoscopy procedures or patient workflows. Future implementation will depend on subsequent studies confirming the cost-effectiveness of integrating these models into medical practice.
The takeaway
Large language models show promise in exceeding the accuracy of traditional random forest machine learning for lesion classification. Watch for upcoming research studies that move beyond accuracy benchmarks to analyze the specific economic and clinical implementation costs.
Further reading
For broader trends in medical diagnostic automation, explore our section on Artificial Intelligence.
Source note: This article includes information reported by Firstwordpharma.
Live Poll
Do you trust artificial intelligence to accurately report results from medical diagnostic procedures?
