Multi-agent LLM screening finds missed rare-disease candidates inside the EHR

Rare disease diagnosis often stalls long before a patient ever reaches genomic testing. In this preprint, investigators built RESCUE, an end-to-end, multi-agent large language model workflow designed to scan the electronic health record for patients who may warrant exome or genome sequencing. The system was built around a practical problem that has limited earlier tools: many screening approaches depend on pre-identified patients or local rule sets that are difficult to reproduce, labor-intensive to deploy, and disconnected from real-world clinical operations.

The clinical target is straightforward. Patients with rare disease frequently travel through a long diagnostic odyssey, and the window for timely genomic assessment can close before anyone recognizes the pattern. RESCUE was designed to push that recognition upstream by combining rule-based logic with machine learning rather than forcing one method to do all the work.

Four coordinated agents split the screening problem into separate tasks

The platform runs through four specialized agents in sequence. The Ontology Agent classifies clinical data into a four-tier genetic-evidence taxonomy. The Modeling Agent then builds a positive-unlabeled XGBoost classifier. After that, the Screening Agent applies the model across the EHR and also flags patients using a rule-based algorithm. The final Review Agent samples clinical notes to judge medical necessity and operational feasibility.

Breaking the workflow into modular agents is not just a software design choice. It mirrors the way human genetics services actually function: one layer organizes evidence, another estimates probability, and a final layer checks whether a candidate makes sense at the bedside and within existing operations. In rare disease, where phenotypes are incomplete and documentation is uneven, that separation may matter more than any single model score.

Holdout performance was solid, and the real-world flagging rate was meaningful

In a retrospective evaluation at a pediatric hospital, the machine-learning model achieved strong discrimination on a holdout set of 12,591 patients, with an AUC of 0.808. That level of separation suggests the system can distinguish higher- from lower-probability candidates better than chance by a clinically relevant margin, even if it does not solve the full diagnostic problem on its own.

The operational readout is more striking. Among 175,842 patients who met screening criteria, those flagged by RESCUE were 7.4-fold more likely to receive subsequent genomic assessments than controls. In practice, that means the system did more than generate an abstract ranking. It identified a subgroup that went on to cross the threshold into genomic workup at a much higher rate than comparable patients who were not flagged.

A diagnostic system that never moves a patient into testing has little value at the bedside.

Chart review suggested the system recovered patients who had been missed

Blinded chart review confirmed previously missed, medically appropriate candidates for exome or genome sequencing, with 80% precision after accounting for prior testing. That result is important because precision in this setting is not simply a machine-learning metric; it reflects whether the signal corresponds to real clinical need. In rare-disease care, a false-positive flag can create noise, but a false-negative can preserve years of delay.

The review step also gives the workflow a practical clinical anchor. Rather than assuming that algorithmic output alone should trigger action, RESCUE uses note review to test whether the alert fits the medical story and whether the institution can act on it. This is the kind of workflow that can reduce friction in an EHR environment where thousands of patients may carry partial clues but only a small fraction are ready for genetics referral or sequencing.

The design also reflects an important biological reality: rare disease often announces itself through scattered fragments, not a single clean diagnostic code. A child may have recurrent hospital encounters, developmental concerns, subspecialty notes, or a family history buried in free text. Machine learning can assemble those fragments into a coherent risk signal, while rule-based logic can preserve transparency and operational simplicity where hard thresholds still matter.

Rule-based logic and machine learning were treated as complements, not competitors

The authors frame RESCUE as a flexible, scalable system in which machine-learning and rule-based approaches complement one another. That framing is clinically sensible. A positive-unlabeled classifier can surface patterns that rigid rules miss, but deterministic rules remain useful when institutions need interpretable triggers, auditability, or rapid local deployment without extensive model retraining.

The preprint also emphasizes local implementation. Rather than being limited to a single pre-labeled cohort, RESCUE is intended to automate deployment across the full EHR at an institution. That matters because rare-disease screening tools often fail at the last mile: even strong models can stall when they need custom curation, manual rule translation, or site-specific coding logic before they can be used in routine care.

Reproducibility and institutional fit remain the central questions

The work was approved by the Boston Children’s Hospital institutional review board, and the data and code are available online through the project repository. Funding came from the National Human Genome Research Institute. Those details support transparency, but they do not change the core scientific question raised by the preprint: how well will a system like this transfer outside one pediatric hospital, where documentation habits, patient mix, and genetics workflows may differ?

Because the study is a preprint, the findings have not yet been peer reviewed. The evaluation was retrospective, and the abstract does not provide long-term patient outcomes, downstream diagnostic yield after testing, or the burden of false positives on clinical services. The model’s holdout AUC and the chart-review precision are encouraging, but they do not establish whether the tool shortens time to diagnosis, increases confirmed molecular diagnoses, or improves family experience at scale.

For now, RESCUE reads as a carefully engineered attempt to make rare-disease screening more operationally realistic inside the EHR. Its strongest claim is not that one model replaces clinical genetics expertise, but that an agentic workflow can help institutions find patients who were already in front of them and still went unrecognized.


Reporting by Dr. A. Nikhil Krishnan, MBBS for Medicine World Foundation, with additional data provided by medRxiv (Clinical Preprints).