July 5, 2023
Can artificial intelligence assess histologic inflammation and predict outcomes in ulcerative colitis?
These early results show great potential for artificial intelligence in interpreting histopathology among patients with established inflammatory bowel disease.
Iacucci M, Parigi TL, Del Amor R, et al. Artificial Intelligence Enabled Histological Prediction of Remission or Activity and Clinical Outcomes in Ulcerative Colitis. Gastroenterology. 2023;164(7):1180–88.e2; https://doi.org/10.1053/j.gastro.2023
In light of a high rate of interobserver variability when pathologists assess histologic inflammation in ulcerative colitis (UC), researchers developed an artificial intelligence (AI)-aided system to evaluate biopsies from UC patients and predict 12-month outcomes.
They trained and tested the system using 535 biopsies from 273 UC patients recruited from 10 sites internationally. Endoscopies were performed using high-definition instruments. Endoscopic activity was graded using the Mayo endoscopic score and the UC endoscopic index of severity (UCEIS). Digital biopsies were graded according to the PICaSSO Histologic Remission Index (PHRI), the Robarts Histological Index (RHI) and the Nancy Histological Index (NHI). Biopsies were graded independently by three expert pathologists. Human pathologists classified 62% to 76% of patients as being in histologic remission at baseline, depending on the scoring system they used.
A convolutional neural network was trained to distinguish remission from activity on a subset of 118 biopsies, then calibrated on 42 and tested on 375. In the testing cohort, compared to human pathologists, the AI system was 87% accurate in distinguishing histologic remission from disease activity as defined by the PHRI, 80% accurate in doing so using RHI criteria, and 81% accurate when human pathologists employed NHI criteria.
In an external validation cohort of 154 biopsies from 58 separate patients enrolled in the study, the system was 92% sensitive and 81% specific in distinguishing histologic activity from remission when using the PHRI criteria. The system produced results within 9.8 seconds on average per slide.
AI predictions of 12-month disease activity based on histology findings accurately corresponded with actual clinical outcomes at one year in 79% to 82% of cases using UCEIS and PICaSSO criteria, respectively. Moreover, the prognostic accuracy of these predictions strongly correlated with human pathologist predictions based on histologic findings.
Study Design: Prospective cohort
Funding: National Institute for Health and Care Research
Allocation: Not applicable
Setting: Multicenter
Level of Evidence: 2b