Retrieval-Augmented Diagnosis (RAD)

• Retrieval-Augmented Diagnosis (RAD) is a diagnostic framework that integrates externally retrieved evidence, such as historical cases and guidelines, to support and refine decision-making.
• RAD systems employ specialized retrieval substrates and adaptive control mechanisms to select, synthesize, and optimize evidence for varied diagnostic tasks.
• Empirical evaluations in radiology, retinal, and psychiatric applications demonstrate significant improvements in sensitivity, accuracy, and interpretability over conventional methods.

Retrieval-Augmented Diagnosis (RAD) denotes a family of diagnostic systems in which inference is conditioned on retrieved external evidence rather than relying only on parametric model memory. In the recent literature, the retrieved evidence can be historical cases, disease-centered guideline summaries, literature and case reports, Electronic Health Records (EHRs), phenotype resources such as HPO and OMIM, DSM-5 criteria, or multimodal image–text archives; the downstream task can be disease classification, differential diagnosis, diagnostic dialogue, radiology impression drafting, or other evidence-grounded decision support (Li et al., 24 Sep 2025, Samanta, 18 Mar 2026, Xia et al., 22 Jun 2026, Kim et al., 6 Nov 2025, Zhao et al., 6 Feb 2025, Chen et al., 10 Apr 2025). Across these settings, the defining property is not merely retrieval, but retrieval that is tied to diagnostic reasoning, evidence selection, and explicit support for the final output.

1. Conceptual definition and relation to retrieval-augmented generation

RAD is best understood as a diagnostic specialization of retrieval-augmented generation. Standard RAG in NLP typically encodes a textual query, retrieves passages from a text corpus, and conditions generation on those passages. RAD adapts this pattern to diagnosis by changing both the query object and the retrieved object: queries may be multimodal, longitudinal, or dialogue-based, and retrieved items are often whole cases, disease guidelines, or diagnosis-oriented evidence bundles rather than generic passages (Li et al., 24 Sep 2025).

Several works make this specialization explicit. In chest radiography, a new chest radiograph is used to retrieve historically similar cases and associated impression text, and the system drafts an evidence-grounded impression rather than a full report; the authors describe this as case-based reasoning and evidence-grounded language generation, motivated by hallucinations, lack of grounding, and safety concerns in purely generative report systems (Samanta, 18 Mar 2026). In rare retinal disease, Evo-RAD defines retrieval-augmented diagnosis as support-set construction around a test fundus image, then turns retrieval itself into a sequential decision process that edits the reference set until it is diagnostically coherent (Xia et al., 22 Jun 2026). In rare brain MRI reasoning, RADAR retrieves Radiopaedia articles and cases to refine candidate diagnoses generated from an MRI caption and clinical history, thereby operationalizing the clinical workflow of consulting literature for unfamiliar findings (Kim et al., 6 Nov 2025).

Other systems broaden the notion of retrieved evidence. MedRAG retrieves similar EHRs and a hierarchical diagnostic knowledge graph encoding critical differences among diseases with overlapping manifestations, then asks an LLM to reason over both (Zhao et al., 6 Feb 2025). MoodAngels retrieves DSM-5 criteria and historical psychiatric cases, and uses those retrievals inside a multi-agent diagnostic and debate framework for mood-disorder classification (Xiao et al., 4 Jun 2025). MRD-RAG treats multi-round consultation as the target object, retrieving disease-centric diagnostic texts and then using an intermediate Analyzer to reason over interconnections among candidate diseases before a Doctor agent asks the next question or provides the diagnosis (Chen et al., 10 Apr 2025).

A recurrent distinction is that RAD is not identical to generic knowledge injection. The multimodal clinical framework titled “RAD” retrieves disease-centered knowledge offline, refines it into concise guidelines, and injects it directly into downstream multimodal diagnosis via contrastive alignment and decoder queries rather than through query-time prompting (Li et al., 24 Sep 2025). This places RAD alongside, but not reducible to, conventional prompt-based RAG.

2. Retrieval substrates and knowledge organization

The retrieval substrate in RAD is highly task-specific. One dominant pattern is the case archive. In radiology impression drafting, the retrieval unit is a single chest radiograph paired with its IMPRESSION section, stored as fused CLIP image and text embeddings in a FAISS inner-product index; the pilot database contains 2,696 such cases (Samanta, 18 Mar 2026). In retinal disease, the initial reference set is the top-$K$ training images returned by a frozen retinal foundation model, and the evolved support set remains a set of labeled retinal images (Xia et al., 22 Jun 2026). In psychiatry, the datastore is a collection of structured medical records and scale profiles, and retrieval is performed over both record text and scale-derived text using BGE-M3 (Xiao et al., 4 Jun 2025).

A second pattern is the disease knowledge base. The multimodal RAD framework for diagnosis constructs a four-source corpus—Wiki, Research, Guideline, and Book—then retrieves the top-$k$ documents per disease label with MedCPT and summarizes them into diagnosis-oriented guidelines $g_i$ using Qwen2.5-72B; these guidelines become both prototypes for representation learning and query tokens for multimodal fusion (Li et al., 24 Sep 2025). MedRAG builds a four-tier hierarchical diagnostic knowledge graph with broad categories, manifestation-based subcategories, disease nodes, and feature nodes, including LLM-augmented distinguishing manifestations; this graph is then searched to produce a disease-specific “diagnostic differences KG” (Zhao et al., 6 Feb 2025).

A third pattern is the literature or ontology corpus. RADAR indexes Radiopaedia articles and cases with all-MiniLM-L6-v2 and FAISS, retrieving top-5 chunks per generated diagnostic question (Kim et al., 6 Nov 2025). The rare-disease gene prioritization framework based on CoT and RAG uses HPO and OMIM, a 64,068-chunk combined knowledge base, PubMedBERT-based dense retrieval, and ColBERT-v2 re-ranking, with HPO contributing shorter, more phenotype-oriented chunks and OMIM contributing longer narrative case descriptions (Wu et al., 15 Mar 2025). MRD-RAG constructs DI-Tree knowledge bases for both modern medicine and TCM, then adds an auxiliary pseudo-medical-history index so that patient-style language is aligned with disease-centric diagnostic text (Chen et al., 10 Apr 2025).

A fourth pattern is the adaptive retrieval substrate for long EMRs. FIND segments an EMR into sentences, labels each sentence as critical, supportive, or irrelevant, computes an information-density score, and activates retrieval only when the record is not already sufficiently informative for direct diagnosis; retrieval itself proceeds by sentence-to-chunk matching, chunk-to-document aggregation, and then differential-diagnosis filtering of the retrieved documents (Jia et al., 20 Feb 2025).

These designs share a common principle: the retrieval object is shaped to match the diagnostic act. In some systems the object is a prior patient; in others it is a guideline prototype, a literature chunk, a disease–feature subgraph, or a pseudo-clinical narrative. This suggests that RAD is less a single algorithm than an evidence-engineering paradigm.

3. Reasoning and control mechanisms

RAD systems differ most sharply in how they transform retrieved evidence into a diagnostic decision. One family uses evidence-constrained synthesis. In radiology, retrieved impression texts are converted into evidence snippets, and the drafting component is required to include explicit case markers such as [Case 1] and [Case 2]; if the output is malformed or insufficiently grounded, a deterministic fallback summarizer concatenates or merges retrieved snippets so that the draft is strictly subsumed by retrieved evidence (Samanta, 18 Mar 2026).

A second family uses support-set optimization. Evo-RAD formulates retrieval as an MDP $$\mathcal{M}=\langle\mathcal{S},\mathcal{A},\mathcal{T},\mathcal{R}\rangle$$, where the state is the current reference set and its graph, the actions are DELETE, INSERT, and TERMINATE, and the reward combines diagnostic correctness, label purity, and semantic density. The final diagnosis is majority vote on the evolved set $\mathcal{X}_T$, and policy learning uses Group Relative Policy Optimization rather than actor–critic value estimation (Xia et al., 22 Jun 2026). Here, retrieval is not a preprocessing step but the object of optimization.

A third family uses question-centric agentic reasoning. RADAR first generates 10 candidate diagnoses from MRI caption plus clinical data, then uses a retrieval agent to produce question–keyword pairs, retrieves Radiopaedia articles and cases, answers the questions strictly from retrieved evidence at low temperature, and finally asks a doctor agent to produce one primary diagnosis and four differential diagnoses (Kim et al., 6 Nov 2025). MedRAG likewise separates retrieval from reasoning, but grounds the latter in a differential-diagnosis KG and retrieved EHR cases; it also computes discriminability scores $\sigma(e_{L4d_i}) = (n-1)/\deg(e_{L4d_i})$ to generate follow-up questions about the most distinguishing manifestations (Zhao et al., 6 Feb 2025).

A fourth family uses adaptive retrieval control. FIND defines information density

$$I(\mathcal{Q})=\sum_{i=1}^n \left(\alpha\,\mathbb{I}(l_i=\mathsf{A})+\beta\,\mathbb{I}(l_i=\mathsf{B})+\gamma\,\mathbb{I}(l_i=\mathsf{C})\right),$$

with $\alpha=1.0$, $\beta=0.5$, and $\gamma=0.1$, then normalizes it and compares the result to $k$0 and $k$1 to decide whether retrieval is unnecessary, necessary, or necessary with an insufficiency warning (Jia et al., 20 Feb 2025). MRD-RAG uses a different controller: at each round it first decides whether a fresh retrieval is needed, then retrieves candidate diseases, runs an Analyzer to compare interconnections and differences, and finally lets a Doctor agent ask focused questions or provide the diagnosis (Chen et al., 10 Apr 2025).

A fifth family uses multi-agent verification. MoodAngels deploys three diagnostic agents—Angel.R, Angel.D, and Angel.C—with no similar-case retrieval, raw similar-case retrieval, and analytic similar-case retrieval, respectively. Their outputs are then adjudicated by a Judge agent, and if disagreement remains, Positive and Negative agents debate whether the patient has a mood disorder before the Judge issues the final binary decision (Xiao et al., 4 Jun 2025).

Across these systems, the control problem is central. Retrieval can be static, dynamic, gated, debated, or citation-constrained, but in each case the system is designed to mediate between too little evidence and too much irrelevant evidence.

4. Evaluation paradigms and empirical performance

RAD research has developed evaluation protocols that go beyond end-task accuracy. In radiology impression drafting, the multimodal fusion retriever improves clinically relevant retrieval substantially over image-only retrieval: image-only Recall@5 is 0.633, fusion with $k$2 achieves Recall@1 $k$3, Recall@5 $k$4, and Recall@10 $k$5, with a best fusion setting reporting Recall@5 $k$6 (Samanta, 18 Mar 2026). The same work reports average top-1 retrieval similarity of about 0.980, average citation coverage of 0.867, and refusal rate 0.000 on its in-distribution test set (Samanta, 18 Mar 2026).

In rare retinal disease, Evo-RAD reports large gains precisely where static retrieval and foundation models are weakest. On Rare-20, the best linear-probed foundation model, RetiZero LP, has sensitivity 13.05%, whereas Evo-RAD reaches $k$7% sensitivity with $k$8% accuracy and $k$9% macro F1; on Retina-31, Evo-RAD reaches $g_i$0% sensitivity and $g_i$1% accuracy (Xia et al., 22 Jun 2026). The paper summarizes this as +21.04% sensitivity over retinal foundation models and +3.56% over retrieval-based and PEFT methods (Xia et al., 22 Jun 2026).

In rare brain MRI reasoning, RADAR improves both Top-1 and Top-5 diagnosis across multiple LLM backbones. With GPT-4o, RADAR reaches Top-1 accuracy $g_i$2 and Top-5 accuracy $g_i$3; the largest absolute Top-5 improvement over a single-agent baseline is +10.19 for DeepSeek-R1-70B, and the paper highlights an “up to 10.2% performance gain” (Kim et al., 6 Nov 2025). In psychiatry, Angel.R already improves accuracy from 0.797 for a GPT-4o baseline to 0.920 on real-world cases, while multi-Angels reaches 0.925 with sensitivity 0.881 and MCC 0.834 (Xiao et al., 4 Jun 2025).

In multimodal classification, the guideline-centered RAD framework reports state-of-the-art performance across four datasets with different anatomies. On MIMIC-ICD53, RAD raises the overall average metric from 54.19 for the best baseline to 57.28, with F1 improving from 36.32 to 39.71; on SkinCAP, the average rises from 86.15 to 88.64; on NACC, RAD reaches 58.12 vs 55.77 for the best baseline (Li et al., 24 Sep 2025). FIND shows a complementary pattern on Chinese EMR diagnosis: on CMEMR it reaches F1 $g_i$4, on ClinicalBench 40.19, and on CMB-Clin 55.38, outperforming direct prompting, Chain-of-Thought baselines, vanilla RAG, and several adaptive RAG baselines (Jia et al., 20 Feb 2025).

In multi-round diagnostic dialogue, MRD-RAG improves the average GPT-based diagnosis score by 0.47 over the same LLM without RAG and by 0.30 over single-round RAG; PMH-based retrieval outperforms retrieval over raw diagnosis text, and human doctors report a +21.75% win advantage for MRD-RAG-PMH over no-RAG LLMs and +18% over single-round RAG (Chen et al., 10 Apr 2025). In disease QA and decision support, MedRAG reaches 66.04% at the most specific $g_i$5 level on the chronic pain dataset, compared with 54.72% for the best baseline, and its proactive questioning experiments show $g_i$6 accuracy rising from 52.83% under 100% masking of key manifestations to 66.04% when no such information is masked (Zhao et al., 6 Feb 2025).

Taken together, these results show that RAD is not a single benchmark regime. Retrieval quality, support-set coherence, downstream label specificity, dialogue informativeness, and grounding behavior are all treated as first-class evaluation targets.

5. Interpretability, grounding, and safety

Interpretability in RAD is usually operationalized as traceable evidence use. In radiology, generated impression sentences are tied to retrieved cases through explicit markers such as [Case 1], and the top-$g_i$7 similar cases remain inspectable; the system also uses confidence-based refusal, abstaining when the top-1 similarity $g_i$8 falls below a threshold $g_i$9, which yields low similarity and refusal on out-of-domain images such as non-CXR photos (Samanta, 18 Mar 2026). In Evo-RAD, interpretability is attached to the sequence of DELETE, INSERT, and TERMINATE actions and to scalar summaries such as purity and semantic density of the support set (Xia et al., 22 Jun 2026).

A more formal notion of grounding appears in the multimodal guideline-centered RAD framework. It introduces textual Guideline Recall, computed from decoder attention over guideline-indicator tokens, and visual grounding via mIoU against lesion annotations. On MIMIC-ICD53, total Guideline Recall rises from 24.76% to 65.62%, with large gains on specific indicators such as ALT and AST; on ChestX-Det10, Avg-D mIoU rises from 15.98 to 19.72 and Avg-P from 17.78 to 22.04 (Li et al., 24 Sep 2025). RADAR emphasizes literature-grounded explanations by forcing the RAG agent to answer diagnostic questions based solely on retrieved Radiopaedia chunks, then using those answers to refine the final diagnostic ranking (Kim et al., 6 Nov 2025).

Grounding, however, is not guaranteed by superficial success. RAG-X isolates this issue in medical question answering by separating retrieval success from generator adherence. Using its relevance function $$\mathcal{M}=\langle\mathcal{S},\mathcal{A},\mathcal{T},\mathcal{R}\rangle$$0 and Context Utilization Efficiency quadrants, it identifies Effective use, Information blindness, Hallucination / Lucky Guess, and Correct rejection. On its best GuidelineQA pipeline, a 14% gap separates system accuracy from actual evidence-based grounding, and 33.9% of responses are classified as “Lucky Guess” despite the system appearing successful under conventional accuracy metrics (Sivakumar et al., 3 Mar 2026). This directly counters a common misconception that retrieval augmentation automatically yields grounded answers.

Another misconception is that citation or retrieval coverage alone establishes clinical correctness. The radiology drafting paper explicitly notes that citation coverage is a proxy: high citation coverage does not guarantee every statement is clinically correct; it only ensures that statements are textually grounded in historical reports (Samanta, 18 Mar 2026). A plausible implication is that RAD systems require separate validation for retrieval relevance, faithfulness to retrieved evidence, and clinical correctness, rather than treating any one of them as a surrogate for the others.

6. Boundaries, misconceptions, and trajectories

A frequent boundary error is to classify any externally informed diagnostic model as RAD. The rare-disease reasoning model RaDaR is an instructive counterexample: it uses Orphanet, HPO, and public case reports to build a large real-plus-synthetic training corpus and then trains a 32B reasoning LLM with CoT and DPO, but it does not implement an explicit retrieval module at inference time. The paper therefore states that RaDaR is not a Retrieval-Augmented Diagnosis system in the technical RAD sense, although its phenotype-anchored synthetic-data pipeline can be viewed as a training-time analogue of retrieval (Chen et al., 23 Jun 2026).

Another misconception is that RAD must be confined to medicine. The later literature applies the same evidence-centric pattern to battery energy storage fault diagnosis, jailbreak defense, and audio deepfake detection. The BESS assistant combines schema-constrained database retrieval, hybrid text-image retrieval, and multi-agent evidence synthesis for traceable fault diagnosis (Ru et al., 2 Jul 2026). Retrieval-Augmented Defense diagnoses whether an incoming prompt is a jailbreak by retrieving similar attack exemplars, inferring the underlying malicious query and strategy, and thresholding $$\mathcal{M}=\langle\mathcal{S},\mathcal{A},\mathcal{T},\mathcal{R}\rangle$$1 (Yang et al., 22 Aug 2025). Retrieval-Augmented Audio Deepfake Detection retrieves similar bonafide audio samples and contrasts them with the query before classification (Kang et al., 2024). This suggests that the RAD abstraction is evidence-centric and not modality-specific.

Within medicine, the dominant trajectories are already visible. Several papers call for scaling to larger and more diverse archives, extending from image-only or text-only contexts to multimodal and temporal settings, strengthening faithfulness evaluation with domain-specific metrics such as RadGraph or label agreement, and tightening human-in-the-loop workflows (Samanta, 18 Mar 2026, Xia et al., 22 Jun 2026, Kim et al., 6 Nov 2025, Chen et al., 10 Apr 2025). Others emphasize structured decision layers: MedRAG uses a hierarchical diagnostic KG with discriminability-driven questioning (Zhao et al., 6 Feb 2025), while Retrieval Augmented Decision-Making uses ISM and AHP to build a weighted hierarchical decision model $$\mathcal{M}=\langle\mathcal{S},\mathcal{A},\mathcal{T},\mathcal{R}\rangle$$2 and aggregate alternative scores via $$\mathcal{M}=\langle\mathcal{S},\mathcal{A},\mathcal{T},\mathcal{R}\rangle$$3 (Wu et al., 24 May 2025). A plausible implication is that future RAD systems will increasingly combine retrieval, explicit structure, and adaptive control rather than treating retrieval as a single prepend-to-prompt operation.

RAD therefore occupies a precise methodological space. It is broader than case-based reasoning alone, narrower than generic knowledge-augmented LLMs, and increasingly defined by three technical commitments: retrieved evidence must be task-aligned, diagnostic reasoning must expose how that evidence is used, and evaluation must separate apparent correctness from verifiable grounding.