RARE-PHENIX looks for the symptoms that belong on a rare-disease shortlist

A clinical record can contain many observations without offering a concise description useful for investigating a rare disease. Selecting that description requires judgement about which findings matter, how precisely they are described and whether an abnormality was present or specifically ruled out.[4]

RARE-PHENIX, published in npj Digital Medicine on 15 October, connects three tasks: language-model extraction from notes, standardization into Human Phenotype Ontology terms and supervised ranking of the resulting features.[1]

A shared name carries clinical meaning

The Human Phenotype Ontology, or HPO, organizes features of disease so software can compare descriptions by meaning. Its terms come with definitions, synonyms and relationships to other terms. This provides a common vocabulary for matching a patient’s features with knowledge about genes and diseases.[3]

Even a spelling difference needs handling. The HPO’s 2024 update describes British and American variants as synonyms for the same concept. A match should survive the choice between “edema” and “oedema”. The harder task is preserving the distinction between a broad description and a more specific finding.[3]

The project’s clinical guide recommends choosing a term by its definition and checking whether a more specific child term applies. It also distinguishes investigated findings that were absent from findings that were observed. These distinctions carry information that a simple collection of symptom words would lose.[4]

The shortlist is a second problem

The study trained on data from 2,671 patients across 11 Undiagnosed Diseases Network sites. The separate Vanderbilt evaluation used 16,357 notes belonging to 143 patients.[1]

In the authors’ February preprint, a Llama-2-70B configuration reached 0.70 ontology-based similarity against clinician-curated lists, compared with 0.58 for PhenoBERT, at the top 50 terms. PhenoBERT received the same ranking stage for this comparison. This measures semantic agreement between phenotype lists.[2]

A long list has more chances to include a relevant finding. It also asks the next reader to inspect more entries. Ranking addresses where a useful feature appears, while standardization addresses what that feature means. Those jobs remain different even when they sit inside one application.

The HPO guide illustrates the role of judgement with short-sightedness: a severe degree can be more suggestive of an underlying condition than a very mild degree that is common in the general population. The same broad label can therefore carry different diagnostic weight in different patients.[4]

The reported evaluation concerns retrospective agreement with clinician curation. The February preprint leaves effects on diagnostic accuracy, time to diagnosis and clinician workload for prospective evaluation.[2]

For the person reviewing the record, a useful shortlist should make those distinctions easier to inspect. A prominent entry still needs the right meaning and a basis in the patient’s notes. The ranking is valuable when it brings that evidence into view.

Report an error or suggest a correction