# A machine learning model of lamina propria fibrosis in eosinophilic esophagitis for prediction of fibrostenotic disease

## TL;DR

A machine learning model was developed to predict fibrostenotic disease in eosinophilic esophagitis by analyzing lamina propria fibrosis in biopsy samples.

## Contribution

The novel contribution is an AI model that objectively quantifies lamina propria fibrosis and predicts fibrostenosis better than traditional pathology methods.

## Key findings

- AI fibrosis scores correlated strongly with pathologist assessments (Spearman's Rs = 0.64–0.69).
- AI scores predicted fibrostenotic outcomes better than pathologists, even in limited biopsy samples.
- Higher AI scores were linked to rings, strictures, and need for dilatation (p < 0.01).

## Abstract

Eosinophilic esophagitis (EoE) is a chronic immune-mediated disease that can progress to fibrostenotic complications. Lamina propria fibrosis (LPF) plays a critical role in this progression but is difficult to assess reliably in routine biopsies. We aimed to develop and validate an artificial intelligence (AI) model to quantify LPF on hematoxylin and eosin (H&E)-stained slides and to evaluate its ability to predict fibrostenotic disease.

We used a cloud-based platform (Aiforia Inc., Cambridge, MA, USA) to train a supervised AI model to recognize several histological features of EoE, including LPF. Our validation cohort consisted of 213 esophageal biopsy whole-slide images, including 100 adult and 113 pediatric samples with mucosal eosinophilia, which were prospectively evaluated in our anatomic pathology service between 2020 and 2021 using a standardized histological scoring system. AI-based LPF scores were correlated with the development of fibrostenotic disease on subsequent endoscopies after a median follow-up time of 31.4 months.

The AI fibrosis score correlated with pathologist-determined LPF (Spearman's Rs = 0.64–0.69, p < 0.0001) and outperformed pathologists' assessments in predicting fibrostenotic outcomes. Higher AI fibrosis scores were associated with the development of rings, strictures, and need for dilatation on follow-up (p < 0.01), including in cases deemed histologically inadequate by pathologists and in the subgroup without prior strictures. In a Cox Proportional-Hazards model, the AI fibrosis score was an independent predictor of strictures (C-index = 0.73, p = 0.004). Importantly, meaningful predictions were achievable with smaller amounts of lamina propria than traditionally deemed sufficient.

This study demonstrates that AI-based quantification of LPF on routine H&E slides provides an objective and clinically meaningful assessment of fibrosis in EoE. The AI fibrosis score predicts fibrostenotic disease more consistently than conventional pathology evaluation and may improve risk stratification even in limited biopsy samples. Integration of digital pathology tools may enhance histological assessment of fibrosis in EoE and support clinical decision-making.