P0137 - An Automated Machine Learning Assessment of Withdrawal Time Is Accurate and Correlates with Adenoma Detection Rate

Rajesh Keswani, MD, MS, Mozziyar Etemadi, MD, PhD, Florencia Garcia Vicente, MS, Anthony Yang, MD, John E. Pandolfino, MD, Daniel Byrd, MS
Northwestern University Feinberg School of Medicine, Chicago, IL

Introduction: Measuring withdrawal time (WT) is a vital component of colonoscopy quality improvement but is limited in practice. Machine learning (ML) may automate quality metrics measurement and facilitate directed feedback. We aimed to 1) develop an algorithm to measure WT (ML-WT) and 2) assess the correlation of ML-WT with manual WT and adenoma detection rate (ADR).

Methods: Endoscopy procedures at a single academic center were recorded beginning in 3/2018 using a cloud-based video recording system. We developed an algorithm to calculate ML-WT via identification of cecal landmarks (appendiceal orifice, ileum, ileocecal valve, and cecal base). We also calculated “high-quality” ML-WT, defined as frames during withdrawal where the lumen was clearly visible (excluding blurry or “red out” frames). Image processing was handled by an 18-layer Resnet convolutional neural network. The model was trained on predictive tasks for detection on 9,498 images. To ensure a realistic distribution, approximately 30% of the training frames were poor quality. ADR was calculated using a 2-year historical mean. We only included colonoscopists with ≥25 recorded normal (i.e., no biopsy or polypectomy) screening or surveillance colonoscopies. Manual WT was extracted for each corresponding procedure from the electronic health record. We used the Pearson correlation coefficient to assess the relationships between normal colonoscopy WT, ML-WT and ADR.

Results: A total of 16 colonoscopists met inclusion criteria (median historical ADR 39.2%; Interquartile Range [IQR] 35.0%, 44.8%). ML-WT was calculated on a total of 1,823 normal screening and surveillance colonoscopies. Median ML-WT was 12.7 min (IQR 10.9, 14.5). Colonoscopist mean ML-WT very strongly correlated with their historical manually calculated mean WT (r=0.97; Figure 1a). The absolute difference between mean physician ML-WT and WT was 0.86 min (SD 0.7). Overall, 79% of the ML-WT was “high-quality” (frames interpretable for polyp detection). ML-WT moderately correlated with endoscopist historical ADR (r=0.66; Figure 1b). Similarly, high-quality ML-WT moderately correlated with ADR (r=0.69; Figure 1c).

Discussion: We report the development of an accurate automated ML assessment of WT that correlates with ADR in a large cohort of screening colonoscopists. We propose utilizing ML-WT to measure and improve colonoscopy quality in settings where abstraction of quality metrics is infeasible.


Figure: A. Machine learning withdrawal time (ML-WT) strongly correlates with manually calculated withdrawal time (WT)
B. ML-WT moderately correlates with adenoma detection rate
C. High-quality ML-WT (time spent during withdrawal where the colon mucosa can be clearly evaluated) moderately correlates with adenoma detection rate


Disclosures:
Rajesh Keswani: Boston Scientific – Consultant.
Mozziyar Etemadi indicated no relevant financial relationships.
Florencia Garcia Vicente indicated no relevant financial relationships.
Anthony Yang indicated no relevant financial relationships.
John Pandolfino indicated no relevant financial relationships.
Daniel Byrd indicated no relevant financial relationships.


Rajesh Keswani, MD, MS, Mozziyar Etemadi, MD, PhD, Florencia Garcia Vicente, MS, Anthony Yang, MD, John E. Pandolfino, MD, Daniel Byrd, MS. P0137 - An Automated Machine Learning Assessment of Withdrawal Time Is Accurate and Correlates with Adenoma Detection Rate, ACG 2021 Annual Scientific Meeting Abstracts. Las Vegas, Nevada: American College of Gastroenterology.