Clinical Diagnosis of Rare Diseases Using Leaky Noisy-OR Bayesian Networks.

This study presents a probabilistic method for the clinical diagnosis of rare diseases using leaky noisy-OR Bayesian networks automatically constructed from Orphanet and Human Phenotype Ontology data. The resulting model represents diseases and phenotypes as binary variables linked by causal probabilities derived from standardized annotations. Loopy belief propagation enables efficient approximate inference of disease posterior probabilities in large networks containing over 8,000 diseases and 9,000 finding variables. Evaluation on real clinical cases achieves 56.2% Top-3 diagnostic accuracy, in line with the reported performance of leading phenotype-based systems. The proposed framework demonstrates that interpretable and knowledge-grounded probabilistic reasoning can achieve state-of-the-art diagnostic performance for rare diseases while maintaining transparency and reproducibility. Unlike deep learning or ensemble models, it provides explicit causal explanations for each diagnostic hypothesis, enhancing interpretability and trust-worthiness.