Quantify false-positive clinical relationships in the knowledge graph

Quantify the false-positive diagnosis–diagnosis relationships produced by the ontology-grounded knowledge graph used in Synthetic Hospital, thereby assessing the graph’s relationship precision beyond the reported recovery of physician-specified associations.

Background

Synthetic Hospital constructs benchmark labels from an ontology-grounded knowledge graph containing diagnosis–finding and diagnosis–diagnosis relationships. The authors validate the graph against a physician-specified reference set of 119 clinically required diagnosis relationships and report that 111 (93%) are recovered. Eight clinically real relationships are not recovered because they are neither encoded in the ontologies nor visible through shared findings, and these are treated as accepted gaps rather than being recovered by lowering the shared-finding threshold.

The reported validation focuses on recall of expected clinical relationships. The paper explicitly leaves the complementary task of measuring false-positive relationships unresolved, which is important because expanding relationship detection could admit spurious associations and compromise specialty-conditioned relevance labels and other graph-derived benchmark judgments.