P2487 - Validation of Novel Mobile App Using Artificial Intelligence to Capture and Objectively Characterize Stool Images in a Clinical Trial

Award: Presidential Poster Award

Award: Presidential Poster Award

Mark Pimentel, MD¹, Ali Rezaie, MD, MSc(Epi)², Ruchi Mathur, MD³, Jiajing Wang, MSc¹, Asaf Kraus, BA⁴, Ben Neigher, BS⁴, Christine Chang, RN⁵, Mohamad Rashid, MD^3
1Cedars-Sinai Medical Center, Los Angeles, CA; ²Cedars-Sinai Medical Center, Los Anegles, CA; ³Cedars-Sinai Medical Center - MAST Program, Los Angeles, CA; ⁴Dieta Health, Oak Park, CA; ⁵Cedars-Sinai Medical Center - MAST Program, West Hollywood, CA

Introduction: In clinical trials of functional gastrointestinal disorders, subjects are asked to report their stool form, usually relying on the Bristol Stool Scale (BSS) Often they also report an “average” score for a given time period. This subjective method has limitations due to patient self-reporting and recall bias. In addition, the BSS summarizes several stool characteristics into a single 7-point scale which decreases data granularity. A novel mobile application asks users to record a digital image of each bowel movement to catalogue and characterize various stool characteristics via artificial intelligence independent of patients’ perception. The aim of this study is to compare 5 separate stool characteristics between 2 expert gastroenterologists and this new app.

Methods: As part of a double-blind randomized clinical trial of IBS, subjects recorded a digital image of their bowel movements during the 2-week screening phase. The images were taken with the Dieta™ application using subjects’ mobile device and were uploaded to a HIPAA compliant cloud computing system. De-identified images were randomly collected and transferred to the Dieta Stool Annotation Portal for two expert gastroenterologists (MP and AR) to evaluate the images based on 5 characteristics: the Bristol Stool Score (1-7 point scale), along with stool consistency , stool edge fuzziness, fragmentation, and volume on a 0-100 visual analogue scale (VAS). The images were also read for the same parameters by a trained AI algorithm. Results were assessed by Intraclass Correlation Coefficients (ICC). ICC estimates were calculated using SAS 9.4 based on mean-rating (k=2), absolute agreement and two-way random-effects model.

Results: A total of 219 randomly selected stool images were from 14 patients (8 female). The range or scores for each variable was (1-7) for BSS and 0-100 for each of consistency, edge fuzziness, fragmentation and volume. This suggested a group of images covered the full spectrum of the evaluable range for each characteristic. There was a strong agreement between the experts and AI for most of stool characteristics (Table). The agreement with AI was not different from the agreement between experts (MP vs AR).

Discussion: This study demonstrates that compared to 2 expert gastroenterologists, AI can accurately and objectively determine several stool characteristics. These new characteristics may be important to clinical trials looking at the effect of therapy on stool form.



Disclosures:


Mark Pimentel, MD¹, Ali Rezaie, MD, MSc(Epi)², Ruchi Mathur, MD³, Jiajing Wang, MSc¹, Asaf Kraus, BA⁴, Ben Neigher, BS⁴, Christine Chang, RN⁵, Mohamad Rashid, MD³. P2487 - Validation of Novel Mobile App Using Artificial Intelligence to Capture and Objectively Characterize Stool Images in a Clinical Trial, ACG 2021 Annual Scientific Meeting Abstracts. Las Vegas, Nevada: American College of Gastroenterology.