HPOQuest: ask the next phenotype question for rare-disease ranking

A training-free agent that starts from sparse HPO-like phenotypes, keeps a probabilistic disease ranking, and picks informative follow-ups, with large Recall@1/5 gains on four cohorts.

Primary source: arXiv:2609.18431: https://arxiv.org/abs/2609.18431

What’s new: Rare-disease diagnosis is slow because patients first present with incomplete, heterogeneous phenotypes across more than 7,000 known conditions. HPOQuest (arXiv:2609.18431, submitted 16 Sep 2026) is a training-free framework for sequential phenotype acquisition: start from a small set of observed phenotypes, maintain a probabilistic disease ranking, and iteratively choose informative follow-up questions to support clinicians during assessment. Confirmed phenotypes update the ranking; all answers update the candidate question set.

Across four benchmark cohorts, the authors report substantial gains from sparse initial phenotypes: up to 30 percentage points at Recall@1 and 45 percentage points at Recall@5. The claim is not “another classifier,” but that active phenotype acquisition can move the needle when the first evidence is thin.

Why it matters: Rare-disease tools often assume a nearly complete phenotype vector. Clinics do not. An agent that decides what to ask next matches the real interview, and a training-free design is easier to inspect and adapt than a fine-tuned black box. For Bitware, this is medical methods with an evaluation story: ranking lift under deliberately sparse starts.

Caveats: Preprint. Benchmark cohorts are not a prospective clinic trial. Recall@k improvements depend on the initial phenotype sparsity and the disease ontology used (HPO-style setup). “Training-free” still embeds assumptions in the acquisition policy and priors. Follow-up questions must be clinically valid and ethically asked; the system does not replace genetic counseling or confirmatory testing. Treat HPOQuest as evidence that sequential information gain helps rare-disease ranking from limited starts, then validate question quality and safety with clinicians before any bedside pilot.