Irritable Bowel Syndrome (IBS) affects 5–10% of the population and remains a diagnosis of exclusion, based on Rome IV symptom-based criteria, rather than analytical, endoscopic or histological biomarkers. Given clinical similarities with inflammatory bowel disease many patients with IBS-like symptoms—such as diarrhea, constipation or abdominal pain—undergo colonoscopy, that has no macroscopic or histological abnormalities, contributing to diagnostic delay and healthcare burden. The absence of identifiable endoscopic features differentiating Rome IV-confirmed IBS from nonspecific symptom presentations highlights a major unmet clinical need. Deep learning models may detect subtle mucosal patterns beyond human perception, potentially redefining the diagnostic role of endoscopy in functional bowel disorders. In this context, the authors aimed to develop an artificial intelligence model capable of identifying endoscopic features suggestive of IBS during colonoscopy.

A total of 183,543 frames from 242 colonoscopy exams performed in Portugal, Spain and Brazil, using five different colonoscopy systems, were included for model development. Only exams with good bowel preparation (Boston Bowel Preparation Scale ≥6) and no evidence of inflammatory bowel disease were considered. Among these, 129 patients had IBS confirmed according to Rome IV criteria. Images were divided into training, validation and testing sets (70% / 20% / 10%). The model was evaluated using accuracy, precision, recall and F1-score.

In the testing dataset, frames from IBS-confirmed patients were identified with 97.1% accuracy, 91.7% precision, 70.6% recall, and an F1-score of 79.5%. Frames from non-IBS patients were classified with 91.7% accuracy, 94.3% precision, 91.2% recall, and an F1-score of 92.6%.

This proof-of-concept study is the first to demonstrate that deep learning can identify endoscopic patterns associated with IBS, even when colonoscopy appears macroscopically normal. These findings suggest the existence of subtle visual signatures of IBS that may be detectable by AI but remain invisible to the human eye, potentially transforming the diagnostic pathway from a purely symptom-based approach to an image-assisted functional assessment during standard colonoscopy.