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

Area of Science:

• Medical Informatics
• Computational Biology
• Rare Disease Research

Background:

• Clinical diagnosis of rare diseases is challenging due to symptom heterogeneity and limited data.
• Existing diagnostic systems often lack transparency and interpretability.
• Integrating structured knowledge from databases like Orphanet is crucial for improving diagnostic accuracy.

Purpose of the Study:

• To develop a probabilistic framework for automated clinical diagnosis of rare diseases.
• To leverage Bayesian networks for modeling causal relationships between diseases and phenotypes.
• To achieve state-of-the-art diagnostic performance with enhanced interpretability and reproducibility.

Main Methods:

• Utilized leaky noisy-OR Bayesian networks constructed from Orphanet and Human Phenotype Ontology (HPO) data.
• Represented diseases and phenotypes as binary variables with causal probabilities.
• Employed loopy belief propagation for efficient approximate inference in large-scale networks (>8,000 diseases, >9,000 phenotypes).

Main Results:

• Achieved 56.2% Top-3 diagnostic accuracy on real clinical cases.
• Demonstrated performance comparable to leading phenotype-based diagnostic systems.
• The model provided explicit, causal explanations for diagnostic hypotheses.

Conclusions:

• The proposed probabilistic reasoning framework offers a transparent and reproducible approach to rare disease diagnosis.
• Knowledge-grounded Bayesian networks can achieve high diagnostic accuracy while maintaining interpretability.
• This method enhances trustworthiness in clinical decision support for rare conditions.