HPOQuest: A Rare-Disease Diagnostic Agent Using Active Phenotype Acquisition
Kamilia Zaripova, Nassir Navab, Azade Farshad, Annalisa Marsico
cs.AI · cs.LG · q-bio.GN
Quick verdict
Training-free sequential HPO questioning improves sparse rare-disease diagnosis—strong on the small MME cohort, mixed elsewhere, no significance tests reported.
Summary
HPOQuest is a training-free framework for sequential phenotype acquisition in rare-disease diagnosis. Starting from three seed Human Phenotype Ontology (HPO) terms, it maintains a softmax-normalized compatibility ranking over catalogue diseases and iteratively selects follow-up questions from an ontology-constrained candidate set (children/siblings of seeds plus LLM-chosen top-level categories). Its candidate-anchored rarity (CAR) policy scores questions by current disease posterior, within-disease phenotype frequency, catalogue rarity, and a specificity boost; three-way simulated responses (yes/yes-branch/no) expand or prune the question set. Acquired phenotypes feed candidate construction with a rare-phenotype anchor score, web/PubCaseFinder retrieval, and two LLM synthesis passes. On four RareBench cohorts (MME n=40, HMS 88, LIRICAL 370, RAMEDIS 624), HPOQuest lifts MME Recall@1/5 from 5%/5% to 35%/50%, exceeding a full-profile DeepRare reference; on other cohorts gains are smaller and top-1 recall sometimes falls below sparse baselines. Stratification shows gains concentrate in seed-recoverable cases. Main limitations: closed-world response simulation treats unrecorded phenotypes as denied, many hyperparameters without sensitivity analysis, no significance testing, small MME sample, single backbone LLM, and no stated code release.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- HPOQuest enables training-free, sequential, ontology-constrained phenotype acquisition starting from only three seed HPO terms for rare-disease diagnosis.
- On the MME cohort, acquisition improves Recall@1/3/5 from 5.0/5.0/5.0% to 35.0/42.5/50.0%, exceeding the DeepRare reference that uses the complete recorded phenotype profile.
- Acquisition gains concentrate in seed-recoverable cases, with limited improvement for non-recoverable cases across cohorts.
- No single acquisition policy dominates: CAR wins at deeper cutoffs on MME and LIRICAL, while EIG and LLM-guided selection perform best on HMS.
Strengths
Controlled like-for-like evaluation design
All sparse-input methods receive identical three seed phenotypes, budget (B=20), retrieval, diagnoser, and judge, with DeepRare's private similar-case retrieval disabled (Section 3); four acquisition policies (LLM-guided, EIG, CAR, RRF) are compared under one harness in Table 2, an unusually clean ablation for this literature.
Full implementation transparency
Section 3 reports every hyperparameter (9 CAR constants, 5 stopping parameters, candidate-construction and synthesis sizes), the quantized model (4-bit Qwen2.5-32B-Instruct on one GPU), and per-patient runtime (370 s CAR vs 496/611 s DeepRare-3/all), enabling independent reimplementation.
Honest failure-mode analysis
Novel problem framing with ontology-aware responses
Clinically deployable design choices
Weaknesses
Headline gains rest on the smallest cohort
Closed-world response simulator may overstate real-world benefit
No statistical inference anywhere
Hyperparameter opacity and overfitting risk
LLM-judge acceptance unvalidated
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is HPOQuest and how does it work?
HPOQuest is a training-free AI framework for rare-disease diagnosis. It starts from only three patient symptoms coded in the Human Phenotype Ontology (HPO), maintains a probability ranking over thousands of diseases, and repeatedly asks the clinician the most informative next phenotype questions (at most about 20). Confirmed symptoms update the disease ranking, denied ones prune impossible question branches, and the acquired evidence is combined with Orphanet/PubCaseFinder and web evidence by a local language model to produce a ranked differential diagnosis.
How much does HPOQuest improve rare-disease diagnosis?
On the smallest benchmark cohort (MME, 40 patients) it raised Recall@1 from 5% to 35% and Recall@5 from 5% to 50% versus diagnosing from the same three seed symptoms, even exceeding a baseline that used the complete recorded phenotype list. On three larger cohorts (HMS, LIRICAL, RAMEDIS) improvements were smaller and mixed: deeper cutoffs (Recall@3/5) generally improved, but top-1 accuracy was sometimes slightly below the strongest sparse-input baseline. Gains were largest when the seed symptoms already pointed to the right disease family.
Does HPOQuest require training, and does patient data leave the clinic?
No training is needed: the system runs a quantized open-weight model (Qwen2.5-32B-Instruct) locally on a single GPU, taking roughly 370 seconds per patient, so phenotype data need not leave the clinic. It does query external knowledge resources (Orphanet, HPOA, PubCaseFinder, web search) for retrieval and evidence, which contain no individual patient data.
Is the HPOQuest code or data publicly available?
The reviewed arXiv manuscript contains no code or data availability statement; it does disclose all hyperparameters and the full evaluation protocol, so independent reimplementation is feasible but public release is unconfirmed. Readers should check the latest paper version or contact the authors.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Review is based solely on the extracted arXiv v1 text; only Fig. 1's caption is available, so visual/appendix content could not be assessed.
- Table values were parsed from extracted text; minor transcription risk in dense numeric cells cannot be fully excluded.
- No supplementary material or code repository link appears in the extracted text; availability statements may exist elsewhere.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.