Enhancing rare disease detection with deep phenotyping from EHR narratives: evaluation on Jeune syndrome

Abstract

Background

Patients with rare diseases frequently experience misdiagnoses and long diagnostic delays. Accelerating their diagnosis is essential to ensure timely access to appropriate care. Given the increasing availability of EHRs, combining artificial intelligence and deep phenotyping from large-scale clinical databases offers a promising approach to identify undiagnosed patients. This study assesses the impact of improved phenotype extraction on a screening algorithm for Jeune syndrome, a rare ciliopathy characterized by skeletal abnormalities.

Methods

Phenotypes from Jeune syndrome patients and controls were automatically extracted from patient unstructured EHRs relying on two thesauri separately: the standard UMLS Metathesaurus and the UMLS+, an enhanced version incorporating additional terms identified through deep learning. The machine learning pipeline that we designed for classifying patients with renal ciliopathy was adapted for Jeune syndrome detection. The model was trained and tested on both the datasets created using the two phenotyping strategies.

Results

Using UMLS+ strongly improved the classification of patients with Jeune syndrome, increasing the sensitivity from 49 % to 95 % while maintaining a 90 % specificity. The review of a subset of misclassified controls showed that most of them (69 %) had other genetic skeletal disorders, indicating that the model also captured patients who would benefit from referral to a bone disease geneticist.

Conclusion

AI-based screening combined with high-quality deep phenotyping can help reduce diagnostic delay in rare diseases. The completeness and accuracy of phenotyping from EHRs have a strong impact on screening performances.