DOP062 Quantitative histologic features of the inflammatory microenvironment using digital pathology prior to adalimumab treatment allows prediction of response in patients with Ulcerative Colitis.
J Shamshoian, F Najdawi, S Degryse, C Jayson, J Brosnan-Cashman, K Kolahi, A Shrotre, H Guay, F S LarouxAbstract
Background
Ulcerative Colitis (UC) is a chronic, immune-mediated disease. Despite the advent of anti-inflammatory biologics, the degree of response to therapy is unpredictable and achievement of deeper responses is limited to a minority of patients, rendering the choice of therapy complex and empiric. Quantitative histologic features of the inflammatory microenvironment in UC has potential to inform treatment selection. To address this hypothesis, we compared such features derived from machine-learning (ML) models with endoscopic response status (Mayo endoscopic score; MES<2) in patients with UC treated with adalimumab (ADA).
Methods
A retrospective analysis of a subset of patients with UC from phase III clinical trial of ADA (NCT02065622) was performed. A suite of ML models, collectively known as IBD ExploreTM (PathAI, Boston1) previously trained to quantify tissue regions (e.g., epithelium) and cells (e.g., lymphocytes) were deployed on hematoxylin and eosin (H&E)-stained whole slide images (WSI; N=343) of rectum and sigmoid biopsies at ADA baseline (week 0). Human interpretable features (HIFs) were extracted from each WSI and correlated with end-of-maintenance (week 52) region-level endoscopic response status using univariate logistic regression. Multivariable logistic regression predictive models were also developed on a subset of the data (277 WSIs) using 5-fold cross validation. Biopsies from each subject were placed in a single split to minimize data leakage. Predictive accuracy was assessed on a held-out test dataset (66 WSIs).
Results
Of the model-derived HIFs assessed at baseline (N=657), 70 were associated (p<0.05) with end-of-maintenance region level MES response status after Benjamini-Hochberg (BH) false discovery rate correction. Examples of HIFs associated with end-of-maintenance MES response and their associated odds ratio are shown in Table 1. A model predicting region-specific end-of-maintenance endoscopic response (L2 penalized, class balance weighted) achieved a sensitivity and specificity of 0.64 and 0.61, respectively, on WSIs not seen during predictive model development.
Conclusion
ML-derived baseline histologic features were associated with response to ADA in patients with UC. Features capturing the co-localization of neutrophils to the epithelium-lamina propria junction had the strongest association to end-of-maintenance region MES response. ML-derived quantitative histology features have the potential to guide therapeutic decisions in UC that leverage mechanistic rationale derived from histology, contributing to precision medicine in immunology.
References
1IBD Explore is for research use only. Not for use in diagnostic procedures.