The Same Patient, Two Descriptions, Two Diagnoses: Why Clinical LLMs Need Semantic Stability Testing

Most clinical LLM evaluations test whether a model gives the right answer for a given input. Almost none test whether the model gives the same answer when the input is rephrased. A new paper shows that this gap is not just a theoretical concern — it is a patient safety risk.

---

The Problem: Same Patient, Different Words, Different Diagnosis

Consider a patient presenting with chest pain. A nurse writes "pt c/o chest pain." The patient says "my chest hurts." A junior doctor describes it as "retrosternal discomfort." A consultant notes "anginal syndrome." These are all descriptions of the same clinical scenario — but if you feed each one into a clinical LLM, you may get different diagnoses, different recommendations, different risk assessments.

A new paper from researchers at the Indian Institute of Science makes this failure mode precise. They show that current clinical LLMs are not semantically stable: semantically equivalent rephrasings of the same patient scenario produce meaningfully different model outputs. This is not a minor inconsistency. It is a patient safety risk.

---

Why Standard Similarity Metrics Fail

The obvious first question is: can you just check whether two descriptions are "similar" and flag cases where they diverge? The answer is no — not with standard tools.

The paper demonstrates a critical gap between lexical similarity and clinical equivalence. Cosine similarity on sentence embeddings treats "no history of diabetes" and "diabetes" as highly similar — they share most of the same words. But clinically, they are opposite. One means the patient has diabetes. The other means they explicitly do not.

Phrase Pair	Cosine Similarity	Clinical Meaning
"diabetes" vs "no history of diabetes"	High	Opposite
"prior MI" vs "acute MI"	High	Different clinical states
"no evidence of malignancy" vs "malignancy"	High	Opposite

This is not an edge case. Clinical text is full of negation, temporality, and conditional framing. Standard embedding models were not trained to distinguish these clinically critical differences. They capture surface-level word overlap, not medical meaning.

---

What the Paper Introduces

The paper proposes clinically-aware similarity metrics that go beyond cosine distance. These metrics understand:

• Negation — "no diabetes" and "diabetes" are clinically opposite, not similar
• Temporality — "prior history of" and "currently experiencing" describe different clinical states
• Clinical context — the same words can mean different things in different clinical contexts

With these metrics in hand, the researchers can reliably determine when two patient descriptions are genuinely semantically equivalent — and then test whether the model produces equivalent outputs.

The finding is clear: it does not. Rephrasing changes the output more often than it should.

---

Why This Matters for Enterprise Clinical AI

First, standard validation misses this entirely. Most clinical LLM evaluations test on fixed benchmarks with fixed wording. They measure accuracy against gold-standard answers. But they do not test whether the model is consistent across rephrasings of the same input. A model can score well on benchmarks and still be unstable in practice.

Second, the real-world impact is concrete. Clinical notes are written by different people with different training, different shorthand, and different levels of detail. A triage note, a patient's own words, a consultant's summary, and a discharge letter may all describe the same patient differently. If your model's output depends on which version it receives, you have a consistency problem that no amount of benchmark accuracy can fix.

Third, this is a deployment blocker. If a clinical LLM gives different recommendations depending on who wrote the note — nurse, patient, junior doctor, consultant — then the model is not safe to deploy. The variability is not in the patient's condition; it is in the model's sensitivity to phrasing. That is exactly the kind of failure that causes harm in production.

---

The Governance Connection

This paper has an important implication that goes beyond technical model evaluation: you cannot govern what is not consistent.

AI governance frameworks — whether internal policies, regulatory submissions, or audit requirements — assume that a system produces reproducible outputs for equivalent inputs. If the same patient described two different ways gets two different recommendations, what does "the model's recommendation" even mean? How do you audit it? How do you explain it to a regulator? How do you defend it in a clinical incident review?

Semantic stability is not a nice-to-have. It is a prerequisite for every other governance activity: risk assessment, bias evaluation, post-market surveillance, and regulatory defensibility. Without it, the rest of the governance stack is built on sand.

---

What This Means Practically

1. Add semantic stability testing to your validation suite. Before deploying any clinical model, test it with multiple semantically equivalent rephrasings of the same scenarios. Measure output consistency, not just accuracy.
2. Use clinically-aware similarity metrics, not just cosine distance. Standard embeddings will miss clinically critical differences. The paper's metrics provide a starting point, though they will need validation on your specific clinical domain.
3. Monitor for stability in production. Do not just test at deployment time. Track whether model outputs remain consistent as input phrasing varies in real clinical notes over time.
4. Document stability as part of your regulatory evidence. If you are preparing a submission or audit trail, include evidence that the model produces consistent outputs for equivalent inputs. This is what regulators expect, even if they do not yet require it explicitly.

---

Limitations Worth Noting

The paper is a strong contribution, but some limitations are worth acknowledging:

• The clinically-aware metrics are proposed but need broader validation across different clinical specialties, languages, and documentation styles. What counts as "clinically equivalent" may vary across contexts.
• The paper focuses on diagnostic outputs. Semantic stability in other clinical tasks — treatment planning, risk stratification, patient communication — may present different challenges.
• The metrics require clinical domain knowledge to configure. This is not a plug-and-play tool; it requires clinical expertise to set up properly.
• The paper does not address how to fix unstable models, only how to detect instability. Remediation remains an open problem.

---