What question did this study set out to answer?

To assess the capability of a deep learning model for predicting clinical outcomes in patients with moderate to severe ulcerative colitis.

January 23, 2026

P0484Deep Learning Model Utilizing Stool Images to Predict Clinical Outcomes of Advanced Therapies in Patients with Moderate to Severe Ulcerative Colitis

Key Points

To assess the capability of a deep learning model for predicting clinical outcomes in patients with moderate to severe ulcerative colitis.
Patients with moderate to severe ulcerative colitis starting advanced therapies were prospectively enrolled.
Weekly smartphone-captured stool images were collected for three months after treatment initiation.
Kaplan–Meier analysis was conducted to evaluate clinical relapse risk based on DLSUC classifications.
Out of 56 patients, 39.3% experienced clinical relapse over a median follow-up of 9 months.
Patients identified as active by DLSUC had a significantly higher relapse risk than inactive patients (log-rank p < 0.0001).
Fecal calprotectin levels at 3 months distinguished relapse risk effectively (log-rank p < 0.0001), while C-reactive protein did not.

Abstract

Abstract Background Reliable early biomarkers that predict long-term outcomes of advanced therapies in ulcerative colitis (UC) remain an unmet clinical need. We previously developed a deep learning stool image model (DLSUC) capable of predicting endoscopic activity based on smartphone-captured stool photographs. This study evaluated whether DLSUC can serve as an early predictive indicator of clinical outcomes in patients initiating advanced therapies. Methods Patients with moderate to severe UC who started biologic agents or small-molecule therapies were prospectively enrolled. Participants submitted weekly smartphone stool photographs for three months after treatment initiation. Clinical relapse during follow-up was assessed using Kaplan–Meier analysis, comparing relapse risk between groups classified as active versus inactive by DLSUC at 3 months. Results Fifty-six patients were included (mean age 40 ± 14.5 years; 67.9% male). Baseline fecal calprotectin (Fcal) and C-reactive protein (CRP) levels were 1305 mg/kg (median, IQR 521–3343) and 0.5 mg/dL (median, IQR 0.17–2.26), respectively. The median ulcerative colitis endoscopic index of severity (UCEIS) was 5 (IQR 4–6). Infliximab was the most commonly initiated therapy, followed by JAK inhibitors, vedolizumab, ustekinumab, and ozanimod. Over a median follow-up of 9 months, 22 patients (39.3%) experienced clinical relapse. Patients categorized as active by DLSUC using 3-month stool images had a significantly higher relapse risk than inactive patients (log-rank p 0.0001, Fig. 1). Fcal at 3 months also differentiated relapse risk (log-rank p 0.0001), whereas CRP did not (log-rank p 0.05). Conclusion A deep learning model applied to smartphone stool images may provide a practical, noninvasive tool to predict early clinical outcomes in UC patients initiating advanced therapies. This approach could complement established biomarkers and enhance personalized treatment monitoring. Reference: 1. Lee JW, Woo D, Kim KO, et al. Deep Learning Model Using Stool Pictures for Predicting EndoscopicMucosal Inflammation in Patients With Ulcerative Colitis. Am J Gastroenterol 2025;120:213-224. Conflict of interest: Prof. Dr. Kim, Eun Soo: No conflict of interest Lee, Hyun Seok: No conflict of interest Kim, Sung-Kook: No conflict of interest Kim, Kyeong Ok: No conflict of interest Jang, Byung Ik: No conflict of interest Lee, Yoo Jin: No conflict of interest Chung, Yun Jin: No conflict of interest Bae, June Hwa: No conflict of interest Kim, Eun Young: No conflict of interest

Bookmark

Cite This Study

Kim et al. (Thu,) studied this question.

synapsesocial.com/papers/69731047c8125b09b0d2000b https://doi.org/https://doi.org/10.1093/ecco-jcc/jjaf231.665

Bookmark