Rare Disease Mining Agents (RDMA)

• Rare Disease Mining Agents (RDMA) are computational systems that extract, structure, and reason over rare-disease signals in biomedical EHRs.
• They combine retrieval, ontology grounding, multi-agent coordination, and human review to enhance diagnosis, cohort mining, mechanistic modeling, and drug repurposing.
• The modular four-stage pipeline processes clinical texts by extracting entities, verifying context, matching ontology codes, and refining outputs with human oversight.

Rare Disease Mining Agents (RDMA) denotes a family of agentic computational systems for surfacing, structuring, and reasoning over rare-disease signals in biomedical data. In the narrow sense used most explicitly in recent work, RDMA is a four-stage framework that processes free-text electronic health record (EHR) documents and outputs Human Phenotype Ontology (HPO) phenotype codes and Orphanet rare-disease codes for each document. In a broader and increasingly common sense, the term also functions as a design pattern for rare-disease systems that combine retrieval, ontology grounding, long-term memory, multi-agent coordination, and human review across tasks such as differential diagnosis, cohort mining, mechanistic modeling, and drug repurposing (Wu et al., 14 Jul 2025, Carbonari et al., 18 May 2025).

1. Conceptual development and methodological lineage

The contemporary RDMA literature emerged from two converging lines of work. One line focused on rare-disease detection and diagnostic policy under extreme data scarcity; the other focused on ontology-grounded text mining and retrieval-augmented LLM reasoning. A survey of LLMs in rare-disease diagnosis describes the resulting design space as one that spans EHR notes, phenotype ontologies such as HPO, disease resources such as Orphanet and OMIM, retrieval-augmented generation, parameter-efficient tuning, structured questionnaires, and multimodal fusion (Carbonari et al., 18 May 2025).

Earlier rare-disease AI systems supplied several of the methodological ingredients later absorbed into RDMA. A model-based reinforcement learning system for prenatal rare-disease diagnosis formalized sequential symptom inquiry as a high-dimensional partially observed decision problem over 220 ultrasound-detectable symptoms and 81 diseases, using entropy reduction as the stopping criterion; on the global task, its DQN-MC-Bootstrap policy converged to about 40 questions on average, compared with 89 for a Breiman tree and 117 for DQN-MC from scratch (Besson et al., 2018). Semi-supervised rare-disease detection work then addressed class imbalance directly: one GAN-based framework reported a PR-AUC of 34.18% versus 29.04% for logistic regression and 28.95% for a neural-network baseline on a rare inherited blood disorder task, while a sequence-modeling GAN for exocrine pancreatic insufficiency achieved 0.56 PR-AUC on 1.8 million longitudinal claims records (Li et al., 2018, Yu et al., 2019). This suggests that RDMA did not arise ex nihilo from LLM prompting, but from a longer synthesis of ontology-aware NLP, sequential decision support, and imbalanced rare-event detection.

Corpora and annotation resources also played a formative role. The RareDis corpus provides 1,041 disease-centered texts from NORD with annotations for Rare Disease, Disease, Sign, Symptom, and Anaphor entities, alongside six relation types. Its reported inter-annotator agreement is 83.5% exact-match F1 for entities and 81.3% for relations, making it a gold-standard resource for rare-disease information extraction (Martínez-deMiguel et al., 2021). Later prompt-based extraction work showed that fine-tuned BioClinicalBERT achieved overall F1 of 0.689 on rare-disease phenotype extraction, while GPT-3.5-turbo reached 0.472 zero-shot and up to 0.610 in one-shot settings, with relatively strong performance on rare diseases and signs in the one-shot regime (Shyr et al., 2023).

2. Core architecture of EHR-centered RDMA

The canonical RDMA framework operates on a corpus $X$ of free-text EHR documents and produces, for each document $x_i$, a set of HPO phenotype codes $\hat{\Phi}_p(x_i)$ and Orphanet rare-disease codes $\hat{\Phi}_r(x_i)$. Its architecture comprises four agentic stages: Entity Extraction, Entity Verification & Implication, Verified Entity Matching, and Dataset Refinement (Wu et al., 14 Jul 2025).

The pipeline is modular. A chunker splits notes into sentence-level or fixed-window chunks. Retrieval tools provide ontology and auxiliary context through HPO and Orphanet databases indexed via FAISS and MedEmbed embeddings, a lab-event reference-range database, and an abbreviation expansion knowledge base. Separate LLM sub-agents serve as extractor, verifier, implication, matcher, and refiner. Data flow proceeds by retrieving top-$k$ ontology candidates per chunk, proposing unverified entities $E_{uv}$, resolving them through abbreviation detection, direct ontology checks, lab-based implication, and LLM implication classification, then matching verified entities $E_v$ to final ontology codes. A refinement agent compares outputs to historical annotations and flags disagreements for human review (Wu et al., 14 Jul 2025).

A distinctive feature of RDMA is its treatment of scattered and implicit evidence. Sentence-level retrieval pairs each chunk with canonical ontology definitions or synonyms, allowing the model to combine local note context with controlled vocabulary context. Abbreviations are explicitly disambiguated; the framework gives “NPH” as a paradigmatic example, distinguishing “Neutral Protamine Hagedorn” from “Normal Pressure Hydrocephalus” by context. Numerical mentions trigger lab-event processing, so out-of-range values can be converted into implied phenotypes, as in “glucose 285 mg/dL” to “Hyperglycemia.” More general implication classification asks whether an extracted phrase implies a phenotype, and any implied phenotype is re-verified against HPO before matching. The system also maintains a global document-level buffer of extracted codes, enabling cross-sentence aggregation of patterns that no single chunk would fully reveal (Wu et al., 14 Jul 2025).

Published use cases illustrate the intended behavior. In one case, RDMA corrected an original annotation that had mapped “NPH” to Normal Pressure Hydrocephalus; contextual review identified the note as referring instead to Neutral Protamine Hagedorn insulin. In another, “uncontrolled seizures despite multiple AEDs” was matched to Orpha:69895 (Dravet-like syndrome). Elsewhere, “Creatinine 2.5 mg/dL” plus “eGFR 25 mL/min” yielded the implied phenotype HP:0000123 (Elevated serum creatinine), and “portal vein thrombosis” was surfaced as Orpha:299152 (Non-cirrhotic, non-tumoral portal vein thrombosis) after ontology lookup and human review (Wu et al., 14 Jul 2025).

3. Benchmark construction and empirical performance

A major contribution of the RDMA line is the creation of realistic evaluation benchmarks that move beyond idealized case reports and ICD-only labeling. MIMIC-RD was built from MIMIC-IV critical-care notes using the RDMA mining pipeline followed by multi-stage human validation. The construction process began with 1,000 randomly sampled discharged patients whose free-text notes were concatenated. For each patient, two parallel LLM extractions were run: a low-temperature pass at $T = 0.01$ to maximize precision and a high-temperature pass at $T = 0.7$ to improve recall. Each pass asked the model to list candidate rare-disease mentions, verify which candidates corresponded to true Orphanet rare-disease entities, and return standardized Orphanet codes. Phenotypes were then extracted and mapped to HPO, with automatic post-processing to remove phenotype labels that duplicated Orphanet disease codes; fewer than 2.5% of extractions were removed in this step (AlDin et al., 18 Dec 2025).

Human validation was performed by four medical students who independently reviewed each candidate rare-disease mention in context and answered whether the note truly described a patient’s rare disease according to Orphanet definitions. Mean pairwise Cohen’s $\kappa$ was 0.71 and Fleiss’ $x_i$0 was 0.71. Only mentions endorsed by at least three of the four annotators were retained, while 2–2 splits were discarded. The resulting benchmark contains 145 patients, 120 unique Orphanet diseases, 192 confirmed rare-disease occurrences, an average of 1.32 rare diseases per patient, and an average of 127.93 phenotypes per patient. Figure 1 of the study compares the roughly 128 phenotypes per patient with RAMEDIS, MME, HMS, and Lirical, and reports a 10× expansion in phenotypic complexity (AlDin et al., 18 Dec 2025).

MIMIC-RD evaluates rare-disease differential diagnosis by asking a model: “Given this list of observed phenotypes, rank the 10 most likely Orphanet rare diseases.” Performance is measured by Hit@$x_i$1 at both disease level and patient level:

$x_i$2

On the disease-level benchmark of 192 instances, Llama 3.3 70B achieved 20.3/35.9/40.1 for Hit@1/5/10, Qwen 32B 18.2/30.2/37.5, Mistral 24B 13.0/27.6/33.3, OpenBio 70B 9.5/17.9/26.3, and Mixtral 70B 6.8/15.8/22.6. Patient-level results were 3–5 percentage points lower but preserved the same ordering. Even the strongest model missed nearly 60% of true rare diseases within its top-10, representing an approximately 35% drop relative to prior curated case studies (AlDin et al., 18 Dec 2025).

The original RDMA extraction study reports complementary mining metrics. For phenotype extraction, the best RAG-HPO baseline using Mistral 24B obtained precision 0.55, recall 0.61, and F1 0.58, whereas RDMA with Mistral 24B achieved precision 0.63, recall 0.68, and F1 0.65. For rare-disease extraction on human-corrected labels, a RAG-RD baseline reached F1 of approximately 0.32–0.36, while RDMA reached F1 of approximately 0.48–0.59, described as up to 30% absolute F1 improvement. The same work reports that RDMA pipelines run four times faster on a 3090 than local RAG-HPO on an A6000 and reach an approximate 10-fold inference-cost reduction relative to cloud GPU rentals (Wu et al., 14 Jul 2025).

4. Extensions into diagnosis, retrieval, mechanistic modeling, and drug discovery

The RDMA concept has been generalized in several directions. RareAgents introduces a multi-disciplinary team framework for rare diseases built around four modules: a Coordinator or Attending Physician Agent, Specialist Agents drawn from a pool of 41 predefined specialties, a Memory Manager, and a Tool Interface that wraps Phenomizer, LIRICAL, Phenobrain, DrugBank, and DDI-Graph. The framework uses iterative specialist discussion, memory retrieval, and tool feedback before a final decision. On RareBench-Public differential diagnosis, RareAgents with Llama-3.1-70B reported Hit@1/3/10 of 0.5589/0.6867/0.7811 versus 0.4169/0.5815/0.7068 for GPT-4o, with median rank 1.0 versus 2.0. On medication recommendation for MIMIC-IV-Ext-Rare, it reported Jaccard 0.4108, F1 0.5563, DDI 0.0796, and #MED 13.17, outperforming GPT-4o zero-shot CoT on all listed metrics. The same study presents MIMIC-IV-Ext-Rare with 4,760 rare-disease patients and 18,522 admissions (Chen et al., 2024).

Retrieval-augmented diagnostic reasoning has also been adapted to rare-disease imaging. RADAR, designed for rare disease detection in brain MRI, embeds case reports and literature with sentence transformers, indexes them with FAISS, retrieves clinically relevant evidence, and passes the result to a final doctor agent. On the NOVA dataset comprising about 900 scans and 281 distinct rare diseases, RADAR reports up to a 10.2% performance gain, with improvements including +4.46% Top-1 and +6.95% Top-5 for GPT-4o, and larger gains for open-source models such as Qwen3-32B, DeepSeek-R1-70B, and MedGemma-27B (Kim et al., 6 Nov 2025).

Mechanistic modeling constitutes a further extension. AgentODE treats RDMA as a two-loop system in which an LLM proposes candidate ordinary differential equation structures and a tool-augmented inference agent iteratively refines parameter distributions using only population-level summary statistics. In a rare-disease case study on recessive dystrophic epidermolysis bullosa (RDEB), the system was given 231 observations across 46 patients on serum albumin, $x_i$3, hemoglobin, and BMI_rel. It recovered a four-variable relaxation-and-coupling system, reported in-sample RMSE of 0.644 and mean normalized summary discrepancy of approximately 0.21, and was contrasted with a baseline that achieved lower RMSE but degenerate structure. This suggests an RDMA interpretation in which “mining” includes extracting mechanistically plausible latent dynamics from small-cohort summary data rather than only extracting entities from text (Yang et al., 1 Jul 2026).

Drug repurposing and biomedical knowledge-graph mining form another branch of the RDMA landscape. RareAgent reframes rare-disease drug repurposing as a self-evolving adversarial debate among Explorer, Proponent, Skeptic, and Principal Investigator agents over a task-specific evidence graph. On the 100,000 Genomes Project cohort, it reports indication AUPRC 0.438 and AUROC 0.662, improving AUPRC by 18.1% over a reasoning baseline; after two rounds of self-evolution, AUPRC rises to 0.463 and AUROC to 0.750 (Qin et al., 7 Oct 2025). ChatDRex supplies a conversation-driven multi-agent system over the NeDRex knowledge graph, with agents for query routing, network analysis, functional coherence evaluation, literature mining, and hallucination-checked finalization; the same work explicitly describes how this design can be extended to rare-disease scenarios through Orphanet ingestion, ClinVar and HPO augmentation, and parameter tuning for sparse seed sets (Süwer et al., 26 Nov 2025). A related literature-mining pipeline for alkaptonuria used PubTator3 on 168,502 PMIDs to build an extended network with 27,252 nodes and 261,649 weighted relations, plus a high-confidence network with 1,450 nodes and 3,307 edges, recovering known AKU mechanisms while prioritizing inflammation, oxidative stress, bone remodeling, and candidate repurposing agents (Pham et al., 16 Mar 2026).

5. Privacy, synchronization, and human oversight

Privacy preservation is a recurrent RDMA design constraint because rare-disease evidence often resides in identifiable clinical notes, institution-specific records, or sparse cross-hospital case collections. The EHR-centered RDMA framework addresses this directly through local inference. Once models and retrieval indexes are on-premise, no calls to external APIs are required. The phenotype pipeline can run a quantized 4-bit Mistral 24B on a single RTX 3090 with 24 GB VRAM, and the rare-disease pipeline has a similar memory footprint. The study categorizes deployment cost from CPU-only through 1×3090, 1×A6000, and 4×A6000, and reports up to 10× lower inference cost than cloud GPU rentals while remaining “HIPAA-friendly” (Wu et al., 14 Jul 2025).

A more general synchronization architecture appears in Autonomous Agent-orchestrated Digital Twins (AADT), which uses the OpenClaw runtime, a proactive heartbeat scheduler, modular Agent Skills, and a unified RDMDT JSON state store. Two prototype skills are described. “PhenoSkill” monitors caregiver free text, redacts PII, extracts HPO terms using PhenoSnap, and appends timestamped PhenoPackets. “ClinVar Monthly Scan” converts patient variants to SPDI, queries NCBI E-utilities, detects VUS to Pathogenic or Likely Pathogenic reclassifications when review stars are at least 2, and updates the digital twin. Reported latency is typically under 15 minutes for phenotype updates and under 24 hours for variant reclassification. In the prototype, synchronization error is approximately 0% by design, and a variant reinterpretation case achieved alert latency under 2 hours on release (Chen et al., 28 Mar 2026).

Cross-institutional privacy has motivated latent communication schemes. MedLatentDx assumes hospital agents retain private case databases, local retrieval, and local LLM backbones, while transmitting only compact latent KV blocks to a host agent. On CrossRare-Bench, which contains 8,022 cases over 235 OMIM conditions and hospital-level partitions, MedLatentDx with a Llama-3.2-3B host reports accuracy of about 0.73 versus 0.35 for a raw-latent baseline and macro F1 of about 0.68 versus 0.32. Under a frozen-LLM continuation attack, raw-KV communication leaked Token F1 of about 0.60–0.76, whereas MedLatentDx leaked Token F1 of about 0.07, alongside substantial BERTScore reduction. The study therefore frames latent communication as a way to improve cross-hospital diagnosis while reducing reconstructable clinical content (Wang et al., 11 Jun 2026).

Human oversight remains central across these systems. MIMIC-RD retained only rare-disease mentions endorsed by at least three of four annotators (AlDin et al., 18 Dec 2025). RareAgents explicitly proposes human-in-the-loop adjudication of intermediate phenotype extractions and explainability layers to show which phenotypes drove a disease score (Chen et al., 2024). AADT formalizes constrained agency: the agent may monitor and propose updates, but may not enact clinical decisions; all updates are timestamped and surfaced for clinician review (Chen et al., 28 Mar 2026).

6. Limitations, misconceptions, and future directions

A common misconception is that recent LLMs already perform adequately on realistic rare-disease differential diagnosis. MIMIC-RD directly contradicts this. Its error analysis identifies phenotype overload, ambiguous descriptions, sparse pretraining coverage for very-rare conditions, and phenotype–disease frequency mismatch as major failure modes. Patients with successful Hit@10 predictions showed more “very frequent” phenotypes for the true disease, whereas “No-Hit” cases more often exhibited only “occasional” or “very rare” HPO signs. The study also notes systematic mis-ranking for clusters dominated by low-incidence phenotypes, such as ocular albinism’s “iris transillumination” and Pompe disease’s “glycogen-filled vacuoles” (AlDin et al., 18 Dec 2025).

Another misconception is that rare-disease mining can be approximated well by ICD-code proxies alone. MIMIC-RD was motivated precisely by the observation that many rare diseases lack direct mappings to comprehensive resources such as Orphanet, so ICD-based labeling significantly undercounts them (AlDin et al., 18 Dec 2025). Likewise, the EHR-mining RDMA study reports that implicit phenotype recall remains below ideal and that lab-event and implication modules each contribute only about 1–2% F1 in ablation, while many implied phenotypes are still missed (Wu et al., 14 Jul 2025). These findings indicate that ontology mapping, implication handling, and contextual verification are necessary but not sufficient.

Future directions in the literature are consistent across multiple strands. For diagnosis, proposed next steps include multimodal inputs combining labs, imaging, genomics, and vitals; synthetic case augmentation for underrepresented diseases; ontology-guided retrieval over synonyms, variant names, and tissue-specific phenotypes; dynamic prompting agents that ask clarifying questions; clinician adjudication of intermediate outputs; and explainability modules (AlDin et al., 18 Dec 2025, Chen et al., 2024). For mining and extraction, the roadmap includes expanding public annotation sets through semi-supervised or active learning, integrating additional knowledge sources such as UMLS and drug-gene interaction databases, and adopting more advanced in-context learning or light fine-tuning targeted at implication inference and medical reasoning (Wu et al., 14 Jul 2025, Carbonari et al., 18 May 2025). For collaboration, privacy-preserving architectures such as local-first inference, event-sourced digital twins, and latent cross-hospital communication suggest that the future RDMA ecosystem will be distributed rather than monolithic (Chen et al., 28 Mar 2026, Wang et al., 11 Jun 2026).

Taken together, the published record presents RDMA not as a single fixed algorithm but as a technical genre. Its defining commitments are ontology-grounded rare-disease representation, agent-mediated retrieval and reasoning, explicit handling of implicit or fragmented clinical evidence, and operational designs that acknowledge data scarcity, privacy constraints, and the necessity of human validation.