Artificial intelligence assessment of endoscopic severity and extent: A machine learning approach to continuous...

The endoscopy subscore is a therapeutic endpoint in ulcerative colitis trials but is limited by reader variability and inability to fully capture the degree of inflammation across the colon. We developed the artificial intelligence assessment of endoscopic severity and extent (AI-ESe), a continuous, more granular assessment of inflammation throughout the colon to better capture the totality of inflammation, rather than a single score based on the worst lesion.

AI-ESe analyzes endoscopic videos through a multistep process including preprocessing for image quality, endoscope stalling (pausing) detection to mitigate mucosal oversampling, and disease severity assessments, generating an inflammatory heatmap. We aimed to validate the stalling and severity algorithms and analyze the overall output of AI-ESe to measure the construct validity and the heterogeneity of inflammation on holdout endoscopic video datasets.

Stalling detection reduced the mean absolute temporal disagreement in position estimates from 39 seconds with uniform sampling to 23 seconds using this novel model against a human reference standard, which is comparable to the interreader variability of humans (16 seconds). Model assessment of disease severity achieved a quadratic weighted kappa of 0.80. The overall output of AI-ESe correlated with conventional classifications of the endoscopy subscore, while capturing substantial heterogeneity in inflammation severity. Among videos with an endoscopy subscore of 3, the proportion of moderately to severely inflamed mucosa ranged from 17.9% to 100%.

AI-ESe enabled detailed assessment of endoscopic inflammation, capturing heterogeneous burden of disease within conventional endoscopic severity classifications in ulcerative colitis. Granular disease assessments using AI-ESe demonstrate an enhanced approach to disease evaluation.