Abstract
Introduction. Classifying stool consistency in infants is essential in pediatric gastroenterology. The Brussels Infant and Toddler Stool Scale (BITSS) is the visual standard, but it exhibits interobserver variability. Artificial intelligence (AI) offers an objective alternative to reduce subjectivity. Objective. To evaluate the agreement between an AI model and the classification performed by pediatric gastroenterologists using BITSS. Population and methods. An observational, cross-sectional, retrospective study based on a photo database of stool samples from infants under 12 months of age in diapers. We selected 1100 images that met quality criteria; a majority of three gastroenterologists agreed on 896, which we used as the reference set. The images were organized into seven BITSS categories and a grouped version. We trained 20 AI models using an 85/15 split for training and validation. We evaluated performance using exactness, sensitivity, specificity, precision, and F1 score. Inter-specialist agreement was assessed using Fleiss's Kappa, and AI-reference agreement was assessed using weighted Kappa. Results. The specialists showed 24.7% complete agreement and 75.3% partial agreement, with a Fleiss Kappa of 0.24. The best model for seven categories achieved an overall accuracy of 0.56, with Kappa values of 0.59 (linear) and 0.75 (quadratic). In the grouped version, accuracy reached 0.76, with coefficients of 0.69 and 0.78, respectively. Conclusion. The AI showed moderate to substantial agreement with the reference classification, particularly in the grouped categories. The findings support the approach's feasibility and warrant external validation studies.