Multicenter Validation of Video-based Deep Learning to Evaluate Defecation Patterns on 3-dimensional High-definition...

Deep learning technologies have demonstrated the ability to identify dyssynergic defecation for diagnosis of common gastrointestinal motility disorders through nuanced interpretation of 3-dimensional high-definition anal manometry (3D-HDAM). We aimed to validate a deep learning algorithm capable of spatiotemporal analysis of 3D-HDAM in a multicenter setting.

We included 1214 consecutive anorectal manometry studies performed across 3 large health care systems between 2018 and 2022. Deep learning results were compared with expert interpretation according to the London consensus protocol as reference standard. Diagnostic accuracy was assessed using bootstrap sampling to calculate area under the curve (AUC). We used Wilcoxon tests to analyze how well the confidence scores from the deep learning model correlated with the likelihood that experts would assign ambiguous labels in cases where determinations were uncertain. Video-based deep learning features were clustered using Gaussian Mixture Modeling to reveal novel dyssynergia subtypes.

The deep hybrid learning algorithm achieved AUCs of 0.99 (± 0.001 standard deviation), 0.90 ± 0.008, and 0.79 ± 0.003 at Dartmouth Health, Henry Ford Hospital, and University of Pittsburgh Medical Center, respectively, performance comparable or superior to solely deep learning or traditional modeling on every cohort. The algorithm appeared capable of reporting confidence aligned with manual expert interpretation of ambiguity (W = -20.50; P < .001; -1.73; P = .08; and -3.22; P = .001). We further identified 2 novel classes of dyssynergia patterns that may represent clinically relevant phenotypes of dyssynergia.

3D-HDAM combined with video-based deep learning is a useful and clinically relevant technology for evaluating anorectal dyssynergia. Future use cases can be expanded to evaluate other motility disorders and their treatment.