Assessment of colonoscopy skill using machine learning to measure quality: Proof-of-concept and initial validation.

Area of Science:

• Gastroenterology
• Medical Imaging
• Machine Learning

Background:

• Low-quality colonoscopy is linked to increased cancer risk.
• Measuring colonoscopy quality is a significant challenge.
• Automated assessment tools are needed to improve quality control.

Purpose of the Study:

• To develop and validate an automated, interactive assessment of colonoscopy quality (AI-CQ) using machine learning.
• To evaluate the accuracy of AI-CQ in measuring key quality indicators.
• To assess the potential of AI-CQ for provider training and quality improvement.

Main Methods:

• Developed an AI-CQ model using machine learning on colonoscopy images.
• Included metrics like insertion time, withdrawal time, and polyp detection rate.
• Incorporated novel metrics: HQ-WT and WT-PT.
• Pre-trained a vision transformer and finetuned for multi-label classification.
• Externally validated the model on 50 colonoscopies from a second hospital.

Main Results:

• AI-CQ achieved 88% accuracy for cecal intubation identification.
• High correlation between AI-CQ and manual measurements for insertion and withdrawal times.
• AI-CQ polyp detection rate was comparable to manual assessment (47.6% vs 45.5%).
• Correctly identified retroflexion (95.2%) and right colon evaluations (100%).

Conclusions:

• An interactive AI tool can automatically assess colonoscopy quality.
• This AI-CQ system shows promise for identifying providers needing remediation.
• The tool can facilitate targeted training to improve colonoscopy skills and patient outcomes.