RAG Boosts Clinical LLM Accuracy but Safety Studies Are Scarce

The case for retrieval-augmented generation in healthcare starts with an uncomfortable baseline: several prominent general-purpose chatbots performed poorly when asked nephrology literature questions. In a Mayo Clinic study published in 2024, ChatGPT, Bing Chat, and Bard each scored below 40% accuracy compared with in-depth searches.[1] That result does not settle how newer models behave, and it predates GPT-4o. It still captures the operational problem that clinicians recognize immediately: a fluent answer is not the same as a reliable answer.

RAG is the attempt to narrow that gap. Instead of asking a language model to answer from its internal parameters alone, a RAG system first retrieves material from a selected knowledge base, then uses that retrieved evidence to generate an answer. In clinical settings, the important question is not whether that sounds more responsible. It is whether the system retrieves the right sources, uses them correctly, and behaves predictably enough for a clinician or governed workflow to catch failure before it reaches a patient.

The Broad Evidence Favors Accuracy Gains, Not Safety Certainty

The strongest overall signal comes from a PLOS Digital Health systematic review published in June 2025. Across 70 studies, RAG improved GPT-4 accuracy from 73.44% to 79.97%. Smaller models showed larger relative gains: LLaMA2-70B improved from 42.20% to 50.40%.[2] That is a meaningful change, especially in settings where a health system is considering whether retrieval can make a model more usable without immediately moving to the largest commercial system.

But the same review is also the reason to be cautious. Only 3 of the 70 studies assessed safety, only 2 addressed privacy, and 78.9% of datasets were English-only.[2] Those gaps matter because clinical deployment does not fail only by giving a wrong answer on a benchmark. It can fail by exposing protected health information, performing unevenly across patient groups, mishandling local policy, or producing an answer that looks sourced but rests on the wrong retrieved passage.

That combination should shape how clinicians read the field. RAG is not an empty wrapper around a chatbot; the measured accuracy gains are real in the reviewed literature. At the same time, the literature is still much better at answering “Did the model match the expected answer?” than “Can this system be safely maintained inside a clinical workflow?”

The Best Controlled Results Are Clinically Interesting

Almanac, a RAG framework reported in Nature in 2024, gives one of the clearer affirmative examples. It achieved an 18% accuracy gain over ChatGPT across medical specialties and reached 91% accuracy in cardiology compared with 69% for ChatGPT.[3] The cardiology result is not merely a technical flourish. It is the kind of domain-specific improvement that makes retrieval worth attention for clinicians who have watched general-purpose models confidently miss important details.

The radiology contrast media consultation study, published in npj Digital Medicine in July 2025, is even more concrete. In that controlled task, RAG eliminated hallucinations, reducing them from 8% to 0%, with χ² Yates = 6.38 and p = 0.012.[4] It also maintained 2.6-second response times, compared with 4.9 to 7.3 seconds for cloud models.[4] For a clinical team, those numbers are legible: fewer fabricated answers, faster local response, and a narrower consultation domain.

The word “controlled” is doing important work. Contrast media consultation is not the same as an open-ended diagnostic workup, a multi-problem inpatient chart review, or a treatment recommendation that depends on rapidly changing local formularies and patient-specific contraindications. The radiology result supports the claim that RAG can reduce hallucinations in a bounded setting. It does not prove that any RAG implementation will be safe across the messier clinical work that surrounds the query.

Retrieval Quality Is a Clinical Control Point

A RAG system can only ground an answer in what it retrieves. That makes the knowledge base a clinical control point, not a back-office detail. If the corpus contains outdated policies, incomplete guidelines, unreviewed PDFs, or documents that conflict without metadata, the generated answer may look better supported than it is. The model can cite a source and still use the wrong source.

This is where “grounded in evidence” becomes too vague to be reassuring. A deployable clinical system needs decisions about what enters the knowledge base, who approves updates, how retired documents are handled, whether local protocols outrank general literature, and how the system exposes retrieved passages to the human reviewer. Retrieval is not just a model behavior. It is a maintenance obligation.

The PLOS review’s finding that vanilla RAG underperformed direct generation in some clinical prediction tasks is a useful warning.[2] Retrieval can help when the task depends on reference material. It can also distract or degrade performance when the architecture does not fit the data structure or clinical question. An answer-generation task, a guideline consultation task, and a longitudinal EHR prediction task do not ask the same thing of retrieval.

EHR Prediction Requires More Than Generic RAG

Electronic health record data add a different kind of difficulty. The relevant facts are not always in a single note or guideline paragraph. They unfold over time: medications start and stop, labs trend, diagnoses are copied forward, and social or operational context may sit in narrative text. A generic retrieval layer can pull fragments without understanding what changed, what persisted, or what is no longer true.

EHR-RAG, described in a 2026 arXiv preprint, is useful as a sign of where the field is moving. The framework reported a 10.76% average Macro-F1 improvement across four long-horizon clinical prediction tasks using GPT-5, with gains generalizing to Claude-Opus-4.5 and LLaMA-3.1-8B.[5] Because this is a preprint rather than fully peer-reviewed evidence as of July 2026, it should be treated as suggestive, not settled.

Its importance is less that one framework improved one metric and more that the design problem is being stated correctly. Longitudinal prediction needs temporal awareness and reasoning paths suited to structured and unstructured EHR data. A system that retrieves a recent note, a lab sequence, and a medication history has to preserve their order and clinical relationship. Otherwise, retrieval may create a larger pile of context without producing a safer answer.

Human Oversight Has to Be Designed, Not Assumed

Most clinical AI discussions eventually invoke a human in the loop. For RAG systems, that phrase is only useful if the loop has a shape. The reviewer needs to see the retrieved evidence, recognize when it is stale or mismatched, understand the model’s uncertainty, and have a clear way to override, escalate, or report failure. If the system produces a polished paragraph with hidden retrieval steps, the human is supervising a performance rather than inspecting a clinical artifact.

Workflow also changes the safety profile. A RAG assistant used by a specialist for a guideline lookup is not the same as one embedded into order entry, triage, discharge planning, or prior authorization. The same accuracy number can mean different things depending on who sees the output, how much time they have, whether the recommendation is interruptive, and whether downstream staff treat it as advice or instruction.

Local deployment may reduce some privacy exposure, and the radiology contrast media study’s faster local response times make that path attractive.[4] But local control does not automatically solve accuracy, update cadence, access governance, or monitoring. Cloud-based RAG raises different privacy and data-handling questions, particularly where protected health information is involved. The evidence base cited here does not justify treating either approach as categorically safe.

What the Current Evidence Supports

The current evidence supports a narrower and more useful conclusion than either enthusiasm or dismissal. RAG can improve clinical LLM reliability. The improvement is visible in a 70-study review, in Almanac’s specialty performance, and in a controlled radiology consultation task where hallucinations fell to zero.[2][3][4] Those findings are strong enough that health systems should not treat RAG as mere marketing language.

The same evidence does not support treating RAG as a plug-and-play safety layer for clinical decision support. The best results come from bounded tasks, curated sources, and architectures matched to the clinical domain. The weakest part of the literature is exactly where deployment decisions need confidence: safety, privacy, bias, language diversity, and real-world governance.

RAG has moved the accuracy problem forward faster than the safety evidence has matured. That is enough to take it seriously, and not enough to hand it unsupervised clinical authority. The responsible question is no longer whether retrieval can improve a model’s answer. It is whether the knowledge base, workflow, maintenance plan, and human accountability are strong enough for the setting where the answer will be used.

References

1. Mayo Clinic nephrology study. PMC. 2024.
2. PLOS Digital Health systematic review. PLOS Digital Health. June 2025.
3. Almanac RAG framework. Nature. 2024.
4. RAG in radiology contrast media consultation. npj Digital Medicine. July 2025.
5. EHR-RAG framework. arXiv. 2026.