MedCaseAgent drafts rare-disease reports. Put a clinician at every evidence gate

THE SIGNAL IN ONE SENTENCE

A rare-disease case report begins with one patient. That is exactly why the evidence cannot be treated casually. On October 26, a research team introduced MedCaseAgent, an open system for drafting multimodal clinical case reports from structured clinical facts and medical images. The paper also announces MedCase-150K, a large case-report corpus, MedCase-Bench, a temporally held-out evaluation suite, and an open model post-trained from Qwen3.5-9B for the task. The authors describe a hierarchical workflow rather than one giant prompt. A planner organizes the evidence and figures. A writer produces a draft. An editor checks clinical fidelity, citations and structure. Final audit and citation-repair passes return to the retrieved evidence. The public code can query PubMed literature, medical images, ClinGen and a local disease index. That is a sensible shape for a difficult job. Case reports document unusual presentations, unexpected treatment effects, new disease associations and events that may be too rare to appear clearly in a large trial. They are not high-level proof on their own, but they can surface a signal that deserves further study. A badly supported report can send readers toward the wrong mechanism, the wrong diagnosis or a false claim of novelty. The useful question is therefore not whether an AI can make a report sound publishable. The useful question is whether every sentence can survive an evidence gate. Start with the patient record. MedCaseAgent expects an input file containing clinical facts grouped into history, presentation, diagnostics, management and outcome, with images stored alongside those facts. That structure can make a messy story easier to inspect. It does not prove that the facts were extracted correctly, that a timeline is complete or that the images belong to the right patient and event. The first gate should be source integrity. A clinician needs to compare every clinical atom with the underlying record. Dates, doses, test results, image findings and outcomes should retain source locations. Contradictory notes should remain contradictory until a qualified reviewer resolves them. Missing information should stay visibly missing instead of being completed by a model that has seen similar cases. Next comes consent and privacy. Rare cases are easier to re-identify than ordinary records because the combination of disease, age, location, imaging and treatment history may be distinctive. Removing a name is not enough. A publication workflow needs explicit patient consent, institutional rules, a minimum-necessary data review and a record of every external service that received clinical material. The public repository lets users point the workflow at an OpenAI-compatible model provider. That flexibility is practical, but it also changes the privacy boundary. A hospital cannot assume that an API endpoint, local model and hosted model carry the same retention, security, regional and contractual terms. The endpoint belongs in the case receipt. Then comes retrieval. The system can search literature and disease resources while drafting. Retrieval can make a report more useful by connecting a rare presentation to current evidence. It can also import an irrelevant paper, an outdated classification, a nearby disease with a similar name or a source that does not support the sentence attached to it. Every retrieved citation needs three checks. Does the source exist? Does it support the precise claim? Does it apply to this patient and context? Citation repair can help with the first two. It cannot decide the third by itself. A paper about an adult cohort may not apply to a child. A molecular association may not establish causation. A treatment described in one case may not be appropriate for another person with the same diagnosis. Novelty needs its own gate. The system is designed to assess whether a case adds something new. That is a hard search problem, not a decorative sentence for the discussion section. A claim of first report should show the databases searched, search date, terminology, synonyms, language limits and the closest prior cases. If the search is incomplete, the report should say rare in the reviewed literature, not first in history. The authors took one good step in the public preprocessing design: the source article and reference report are kept separate from the agent input. For source-derived cases, preprocessing writes identifiers to a provenance file, and generation can exclude the source from retrieval. That separation matters because a benchmark becomes meaningless if the system retrieves the report it is supposed to recreate. The repository is also candid about the remaining boundary. Its public preprocessing does not reconstruct the full paper dataset or benchmark split by itself. The example exclusion list is empty, and the documentation says no benchmark exclusion list is implied. Users must supply relevant identifiers and evaluation exclusions. That warning should stay attached to every benchmark result. A temporally held-out set can reduce one kind of leakage by placing evaluation cases after the training period. It does not eliminate memorization from pretraining, overlap through secondary sources, retrieval of the target report or contamination in a connected model provider. Each path needs a separate test. The paper reports that MedCaseAgent outperforms evaluated general-purpose and specialized medical systems under rule-based and model-judged measures, with blinded expert review and ablation studies providing additional support. Those results are author-reported in a new preprint. The abstract available at publication time does not provide enough detail to turn them into a clinical claim. More importantly, report quality is not patient outcome. A well-structured manuscript can still contain a wrong interpretation. A high citation score can reward tidy references while missing a dangerous contradiction. An expert reviewer can rate clarity and relevance without validating every source record. Clinical validation would require a defined intended use, representative sites, prospective testing, error analysis and evidence that the workflow improves a real process without creating new harm. For now, the safest role is drafting support inside an accountable research workflow. The clinician should control the input, approve the retrieval plan, inspect the evidence table, resolve contradictions, review every image description, verify each citation and sign the final report. The patient or authorized representative should understand what will be published. The institution should know where data traveled and how long they remain available. The machine should keep a receipt. That receipt can be compact. For each claim, store the patient-source location, supporting external citation, tool call, model version, human reviewer and final disposition. For each image, store acquisition date, modality, original identifier, transformation history, caption source and approval. For each removed detail, record why it was suppressed. For each unresolved uncertainty, show where it appears in the report. The final gate is publication. Before submission, a person who was not responsible for generating the draft should try to break it. Can they trace the central claim? Can they find a contradictory record? Can they identify patient details that survived de-identification? Can they reproduce the novelty search? Can they open every source? Can they tell which language came from the system and which judgment came from the clinical team? If not, the draft goes back through the gates. This may sound slower than pressing generate. Good. The point of medical automation is not to move uncertainty faster. It is to help a qualified team organize difficult evidence without losing the patient, the source or the right to stop. MedCaseAgent makes the workflow visible enough to have that conversation. The next useful milestone is not a prettier report. It is a prospective, privacy-safe study showing that the gates catch errors before a draft becomes part of the medical literature.