S543 Validation of Novel Mobile App Using Artificial Intelligence to Capture and Objectively Characterize Stool Images in a Clinical Trial

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 1). The agreement with AI was not different from the agreement between experts (MP vs AR). Conclusion: 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.Table 1.: Title: Table. Reliability Statistics for Categorical Questions Evaluating Bloating and Distension Symptoms Footnotes: ***, near perfect agreement (k ≥ 0.81); **, substantial agreement (0.61 ≤ k < 0.81); *, moderate agreement (0.41 ≤ k < 0.61); k, Cohen's kappa coefficient; CI, confidence interval; Sx, symptom; d, days; hrs, hours