The claim that was scored
Miotto and colleagues present a preliminary study of hyperbolic graph representation learning for Mendelian-disease differential diagnosis on a patient-integrated biomedical graph. That graph joins ontology-derived hierarchy with transversal links among phenotypes, diseases, genes, proteins and patients.[1]
On isolated ontology subgraphs, hyperbolic models achieve strong performance in substantially lower dimensions than Euclidean baselines. The same models are then evaluated on a link-prediction task that ranks candidate diseases for each patient.[1]
What the score actually is
Link prediction ranks how likely a candidate disease is to attach to a patient node. That is not the same object as a clinician's diagnosis from the same signs, the false-positive burden, or a change in treatment. A dimension win on an ontology subgraph also does not measure whether the patient-level ranking is clinically useful.[1]
The paper is an unreviewed preprint and names itself a preliminary study. There is no external site, no independent lab repeating the same link-prediction protocol on a new patient cohort, and no pre-declared ranking metric beyond the authors' own setup. "Strong performance" remains their evaluation.[1]
The next measurement
On this evidence ladder the result sits at internal validation that ontology geometry can be carried into a patient-node ranking. What would move it is a frozen protocol rerun on new patients, plus calibration of that ranking against a clinician list—not a further cut in dimension on the original site.[1]