Challenges of Ontology-Based Concept Normalization for Deep Phenotyping in Rare Skin Diseases

Extracting and standardizing phenotypic information from free-text medical reports remains a main challenge in biomedical natural language processing (NLP). Conceptual normalization, which maps textual mentions to standardized vocabularies such as the UMLS or HPO, is a key step for deep phenotyping. Dystrophic epidermolysis bullosa (DEB) is a rare severe and heterogeneous skin disorder causing skin and mucosal fragility that leads to early morbidity and mortality. To better understand the disease's natural history and severity spectrum, it is important to comprehensively identify DEB-related phenotypes. This study aimed to assess a pipeline for automatically mapping clinical terms from textual reports to the UMLS Metathesaurus and to evaluate how well the UMLS covers DEB-related phenotypes. The study was conducted at Necker-Enfants Malades Hospital in Paris, a reference center for rare diseases, using its "Dr Warehouse" database, containing more than 11 million clinical documents. It involved 198 patients with dystrophic epidermolysis bullosa (DEB). Phenotypes were automatically extracted from clinical texts using named entity recognition (NER) and then normalized using a cosine similarity method. Among the 198 DEB patients followed at Necker Hospital, 14,734 documents were analyzed, allowing the extraction of 33,347 phenotypes using an NER model. Of these phenotypes, 13,485 were correctly mapped to a UMLS or HPO concept, while 19,862 were not. A manual evaluation of 300 unmapped phenotypes revealed five main causes of failure: overly specific phenotypes (72%), missing synonyms in terminologies (30%), multiple phenotypes within a single extraction (15%), absence of a real phenotype (9%), missing concepts in terminologies (3%), or spelling/grammatical errors (3%). Despite the initial association failure, a significant proportion of overly specific or poorly formulated phenotypes actually corresponded to existing concepts. These results highlight the limitations of current ontologies and the challenges of automatic medical language processing.