Semantic distance for hypothesis generation in differential diagnosis: a framework to support the pivot and cluster strategy

Abstract The pivot and cluster strategy (PCS) is an emerging approach to diagnostic hypothesis generation. In this strategy, clinicians generate an initial intuitive diagnosis (the pivot) along with a cluster of diseases with similar clinical representations for systematic consideration. Although PCS was proposed to prevent diagnostic error, a knowledge gap remains in its practical implementation: existing differential diagnosis lists rely primarily on individual expertise, lacking reproducibility and scalability. We have previously proposed that deliberately exploring diseases at a moderate semantic distance, that is, diseases moderately similar to the pivot, may foster diagnostic creativity. Building on that proposal, this paper specifies it as a rank-defined, auditable research programme, drawing on two established methods: ranking candidate diagnoses by embedding-based similarity, and building disease vectors from clinical presentation, not disease names. Each disease is represented as a disease concept vector and positioned within a high-dimensional semantic space (the disease map). Diseases with similar clinical features are thereby intended to lie close together, a testable design rather than an established property. Candidate diagnoses are ranked by cosine similarity to a given pivot disease. Because training is stochastic and corpus-dependent, the ranking is auditable rather than strictly reproducible. Taking a clinician-generated pivot as input, the framework supports bounded exploration of moderate-distance candidates; it does not generate the pivot. The disease map is offered as a research agenda rather than a validated tool, and we outline diagnosis-referenced evaluation of map-level retrieval validity, followed by prospective testing of its effects on clinicians’ hypothesis generation and diagnostic performance.