PubCaseFinder: A Case-Report-Based, Phenotype-Driven Differential-Diagnosis System for Rare Diseases

Abstract

Recently, to speed up the differential-diagnosis process based on symptoms and signs observed from an affected individual in the diagnosis of rare diseases, researchers have developed and implemented phenotype-driven differential-diagnosis systems. The performance of those systems relies on the quantity and quality of underlying databases of disease-phenotype associations (DPAs). Although such databases are often developed by manual curation, they inherently suffer from limited coverage. To address this problem, we propose a text-mining approach to increase the coverage of DPA databases and consequently improve the performance of differential-diagnosis systems. Our analysis showed that a text-mining approach using one million case reports obtained from PubMed could increase the coverage of manually curated DPAs in Orphanet by 125.6%. We also present PubCaseFinder (see Web Resources), a new phenotype-driven differential-diagnosis system in a freely available web application. By utilizing automatically extracted DPAs from case reports in addition to manually curated DPAs, PubCaseFinder improves the performance of automated differential diagnosis. Moreover, PubCaseFinder helps clinicians search for relevant case reports by using phenotype-based comparisons and confirm the results with detailed contextual information.

Keywords: Human Phenotype Ontology; PubCaseFinder; case report; differential diagnosis; disease-phenotype association; rare disease.

Figure 1

Distribution of the Number of Case Reports Published per Year in PubMed from 1980 to 2017

Figure 2

Process of Identifying Disease-Phenotype Associations from Case Reports The set of titles and abstracts of case reports were annotated with HPO terms and ORDO terms using ConceptMapper with HPO and ORDO (step 1), and annotations with inappropriate synonyms were excluded using the Allie database (step 2). DPAs were identified in processed annotations (step 3).

Figure 3

Overlap between Two Sets of ORDO Terms Found in Disease-Phenotype Associations from Orphanet and from Case Reports

Figure 4

PubCaseFinder at a Glance and Integration of PubCaseFinder with IRUD Exchange and PhenomeCentral Once a user types an affected individual’s phenotype in the search box, PubCaseFinder displays candidate HPO terms. This enables rapid entry of HPO terms because users select appropriate HPO terms from the list (A). The affected individual is then compared with all rare diseases in Orphanet on the basis of phenotypic similarity, and the ranked list of rare diseases is shown (B). The higher the phenotypic similarity, the higher the displayed probability as a candidate disease. Users can also obtain a ranked list of published case reports in the same manner (C). The context in which a DPA appears is useful for confirming detailed contextual information on the presence of DPAs (D). This figure also shows the integration of PubCaseFinder with IRUD Exchange (a customized system of Patient Archive) (E) and PhenomeCentral (F) via the PubCaseFinder application programming interface (API). The PubCaseFinder API is also developed as the Matchmaker Exchange (MME) API.

Figure 5

Performance Evaluation of PubCaseFinder Clinical cases of rare diseases were collected from PhenomeCentral (step 1), and a ranked list of rare diseases based on phenotypic similarity was obtained with PubCaseFinder for each clinical case (step 2). The performance of PubCaseFinder was evaluated via the “recall at ranks” metric (step 3).

Figure 6

Performance Comparison of Three Different Settings of PubCaseFinder Recalls were calculated on the basis of ranked lists of 2,323 rare diseases for 135 clinical cases from PhenomeCentral.

Figure 7

Performance Comparison of PubCaseFinder (using DPAs from Orphanet and Case Reports), Phenomizer, and Orphamizer Recalls were calculated on the basis of ranked lists of 2,323 rare diseases for 135 clinical cases from PhenomeCentral.

Figure 8

Recalls of PubCaseFinder for Rare Diseases Not Included in Disease-Phenotype Associations from Orphanet Recalls were calculated on the basis of the ranked lists of 1,589 rare diseases via PubCaseFinder for 59 clinical cases from PhenomeCentral.

Figure 9

Distribution of Numbers of Disease-Phenotype Associations from Case Reports and Top-10 Recall Rates For each set of DPAs ordered according to the frequency of occurrence in case reports, we counted the number of DPAs and calculated the top-10 recall rate to evaluate the performance of PubCaseFinder from the set of DPAs. Bars indicate case reports, and a solid line indicates top 10 recall rates.