Angela Hsu, MD1, William Karnes, MD2, James Requa, BS3, Andrew Ninh, 3, Tyler Dao, 3 1UC Irvine, Irvine, CA; 2Digestive Health Institute, University of California Irvine Medical Center, Irvine, CA; 3Docbot, Inc., Irvine, CA
Introduction: Artificial Intelligence (AI) is poised to positively impact patient care and clinical workflows. Beyond assisting polyp detection during colonoscopy, artificial intelligence (AI) has the potential to automatically record key quality measures during colonoscopy, including cecal intubation rate and withdrawal time. We previously developed and reported algorithms to detect first frames of insertion, cecum, and withdrawal. Here we report video validation of these algorithms.
Methods: Ninety-six deidentified videos recorded as part of a multicenter randomized trial were utilized. The ground truth was established by video review, marking the times of insertion (gtIN), first clear view of cecum (gtCEC), and withdrawal from the colon (gtOUT), relative to the first frame of the video. Ground truth withdrawal time (gtWT) was calculated by subtracting gtCEC from gtOUT. A secondary comparator included withdrawal times manually recorded by study coordinators in the AIDA trial (manWT). AI algorithms for IN, CECUM and OUT were subsequently run on all videos to record times of aiIN, aiCEC and aiOUT, respectively. We then compared withdrawal times between ground truth, manually entered and AI-predicted values (gtWT vs manWT vs aiWT).
Results: Of the 96 videos reviewed, one did not reach cecum so gtWT could not be calculated. AI did not identify cecum in 2 videos, one of which was a true negative (cecum was not reached) and the other was a false negative. There was one false positive aiCEC identified 16 minutes before cecum was reached. Correlation between aiWT and gtWT was excellent (Figure 1A), and performed much better than manWT (Figure 1B). The mean difference between ground truth and ai-predicted withdrawal time (gtWT-aiWT) was 31 seconds, whereas the mean difference between gtWT and manWT is 130 seconds. This difference was statistically significant (p value of < 0.001). AI identified cecum with 98% sensitivity and specificity.
Discussion: This video validation study demonstrates the feasibility of automated and accurate recording of cecal intubation rate and withdrawal time using artificial intelligence. Its accuracy exceeds manual recording during colonoscopy by study coordinators.
Figure: Figure 1A (left panel) Comparison of ground truth (video review) withdrawal times (gtWT) vs. AI-predicted withdrawal time (aiGT). The outlier represents a single false positive cecum identified by AI 16 minutes before cecum was actually reached.
Figure 1B (right panel) - Comparison of ground truth video review withdrawal time (gtWT) vs manually entered withdrawal time by AIDA trial study coordinators (manWT).
Disclosures: Angela Hsu indicated no relevant financial relationships. William Karnes: Docbot – Consultant, Stockholder/Ownership Interest (excluding diversified mutual funds). James Requa indicated no relevant financial relationships. Andrew Ninh: Docbot, Inc. – Employee, Patent Holder, Stockholder/Ownership Interest (excluding diversified mutual funds). Tyler Dao: Docbot – Employee.
Angela Hsu, MD1, William Karnes, MD2, James Requa, BS3, Andrew Ninh, 3, Tyler Dao, 3. P2225 - Automated Cecal Intubation Rate and Withdrawal Time with Artificial Intelligence. a Video Validation Study, ACG 2021 Annual Scientific Meeting Abstracts. Las Vegas, Nevada: American College of Gastroenterology.
