Inferencing Medical Knowledge from an Artificial Neural Network in Inflammatory Bowel Disease

Abstract

Introduction: Inflammatory bowel diseases (IBD), including Crohn's disease (CD) and ulcerative colitis (UC), pose significant challenges to both clinicians and researchers. The complex interaction of genetic predisposition, environmental influences, and dysregulated immune responses underpins the heterogeneous nature of these diseases. This complexity often leads to diagnostic uncertainties and therapeutic dilemmas. Artificial intelligence (AI), particularly neural networks, offers promising solutions for these challenges. Our study aimed to develop a neural network that integrates multiple non-invasive parameters for monitoring inflammatory status in patients with IBD, assess the network's diagnostic accuracy, and use the training data to classify the clinical relevance of biomarkers. Materials and Methods: The study included patients diagnosed with CD or UC based on endoscopic and histopathological criteria following their consent for colonoscopy. Participants were recruited from the Regional Institute of Gastroenterology and Hepatology Cluj-Napoca, Romania. A subset was enrolled prospectively from 2020 to 2022, and the remainder was identified retrospectively from hospital records spanning 2017 to 2020. Data from 70% of the patients were utilized to train the neural networks, and 30% to validate it. The data used as inputs were clinical scores, C-reactive protein (CRP), erythrocyte sedimentation rate (ESR), and fecal calprotectin (FC). Results: The neural networks showed significant diagnostic capabilities, particularly for endoscopic activity in UC (AUC of 0.955) and CD (AUC of 0.813), though histologic activity prediction was less accurate. Regarding the importance of the markers used for the decision, we found that the Crohn's Disease Activity Index (CDAI) was the most significant predictor for CD endoscopic activity and FC for UC. The study highlights the potential of AI in supporting less experienced endoscopists, reducing observer variability, and minimizing the need for repeated colonoscopies. However, further refinement is needed to improve histologic assessment accuracy.