Healthcare · · 7 min read
How to evaluate clinical AI for overtesting
Acumen Editorial · Updated November 21, 2026
Most clinical AI evaluations ask one question: did the model get the diagnosis right. A new study in NEJM AI asks a second one. What would it cost if we did what the model said?
The answer is about three times what physicians ordered. That matters to every health system moving AI from answering questions to drafting plans, pending orders and suggesting referrals. Accuracy is necessary. It is not the same as appropriate care.
What did the NEJM AI study find?
Aran and colleagues gave 24 AI systems, 20 general-purpose models and four specialized medical systems, the same 200 de-identified outpatient and emergency department notes the treating physicians worked from. Each system wrote its own assessment and plan. Tests and specialist referrals were priced at 2026 Medicare rates. The study was published online October 15, 2026.
The results, as reported by 2 Minute Medicine and TechTarget:
- AI plans averaged $314 per visit in tests and referrals. Physician plans averaged $97, about 3.2 times less.
- The mean difference was $216 per visit (95% CI $177 to $257): $105 from testing and $111 from referrals.
- System averages ranged from $209 to $599 per visit.
- In 133 encounters where physicians ordered no tests, the AI systems added testing in 45 to 79 percent of them, most often routine panels such as a CBC, metabolic panel, HbA1c, TSH and lipids.
- AI plans added 1.14 new medications per case, against 0.59 for physicians.
The researchers said the pattern "parallels human defensive medicine." They also found value: nearly a fifth of AI plans matched guidelines and caught orders the physician had missed. That is the point. Some additions are good care. Many are not. An evaluation has to tell them apart.
The study has limits. It was retrospective, the systems saw written notes rather than patients, and it assumed every recommendation would be carried out. New medications were counted but not priced, and the primary outcome was cost, not patient outcomes. Those limits are a reason to rerun the test on your own encounters with your own clinicians, not a reason to set it aside.
Why don't accuracy benchmarks catch overtesting?
Because they don't score it. Most medical benchmarks reward a correct diagnosis, a complete answer or safe advice. A model that orders a full workup on every patient rarely loses points for being thorough. It often gains them.
Low-value care is expensive at scale. Shrank and colleagues, writing in JAMA in 2019, estimated US spending on overtreatment and low-value care at $75.7 billion to $101.2 billion a year. Choosing Wisely recommendations from the Society of General Internal Medicine advise against routine lab testing during checkups and routine pre-operative testing before low-risk surgery. Those are exactly the reflexive orders the NEJM AI study saw.
A model can be accurate and still push care in the wrong direction. If your evaluation only grades diagnosis, you will not see it until the claims data arrive.
What should an overtesting evaluation measure?
Measure what the model adds, what it costs and whether a clinician would defend it.
| Metric | What it catches | How to score it |
|---|---|---|
| Excess cost per encounter | Overall drift toward more care | Price added tests and referrals against your fee schedule or Medicare rates |
| Add-on rate when the physician ordered nothing | Reflexive workups | Share of no-test encounters where the model adds any test |
| Low-value order rate | Orders guidelines advise against | Map each order to Choosing Wisely or your own stewardship lists |
| Appropriateness of each addition | Good catches vs. noise | Physician panel labels each addition necessary, reasonable or unnecessary |
| Missed necessary care | Undertesting after cost tuning | Physicians flag omitted orders the case required |
| Diagnostic agreement | Accuracy lost to cost controls | Compare diagnoses before and after any cost-focused change |
Cost alone is the wrong target. A model that orders nothing scores well on cost and badly on care. Report excess cost next to appropriateness and missed care, never on its own.
How do you build an overtesting test set?
Use your own encounters, not public notes. The NEJM AI study used a public sample collection of uncertain date and setting. Your order sets, payer mix and patient population are different.
- Sample real encounters across the settings you will deploy in: primary care, urgent care, emergency department, specialty clinic
- Include a deliberate share of encounters where the right answer is to order nothing
- Split each note before the physician's assessment and plan, so the model sees only what the clinician saw
- Keep the documented plan as a comparator, not a gold standard, since physicians can undertest too
- Have two or more specialty-matched physicians adjudicate every added order, with a third to break ties
- Price orders with your own fee schedule or contracted rates, and record the source
- Hold the set out of vendor hands and refresh part of it each quarter
- Rerun the full set after every model, prompt or retrieval change
Does a cost-conscious prompt fix it?
Not by itself. In the NEJM AI study, cost-conscious prompting was tested on 19 systems. It lowered their cost ratios, but 17 of the 19 showed lower agreement with physician diagnoses, per 2 Minute Medicine. Prompting traded one problem for another.
That is why we test cost and accuracy together. Any change meant to cut orders, whether a prompt, a fine-tune or a guardrail, should be rerun against the same physician-graded set. If excess cost drops and missed necessary care rises, the change failed.
Isn't the clinician in the loop enough?
It helps. The NEJM AI authors argued that a clinician who has seen the patient should decide what gets ordered. But review is not a free safeguard. A JMIR study of 20 emergency and trauma physicians at four German hospitals found they acted on 68 percent of AI-generated abstracts, and correctly rejected false ones only 43 percent of the time. Plausible output gets through.
A pre-populated order list is plausible output. Measure overtesting before deployment, then monitor order rates after it, so reviewers are not the only line of defense.
What should you ask a clinical AI vendor about overtesting?
- Have you measured your system's orders and referrals against physician plans on real encounters?
- What is your add-on rate in encounters where the physician ordered nothing?
- Who judged whether added orders were appropriate, and what were their specialties?
- How do you handle cost or stewardship guardrails, and what did they do to diagnostic accuracy?
- Can we run our own private test set against every release before it reaches clinicians?
- What order-rate monitoring do you provide after go-live?
A vendor that only reports benchmark accuracy has not answered the question.
How Acumen helps
We build private, physician-graded evaluation sets for clinical AI. For overtesting, that means real-case tasks drawn from your settings, practicing clinicians who label every added test, referral and prescription as necessary or not, and rubrics that score cost, appropriateness and missed care side by side. We rerun the set with every model release, so you see drift before your claims data do.
See our medical LLM evaluation service, or read how to evaluate a clinical LLM.
FAQ
What is clinical AI overtesting?
Clinical AI overtesting is when an AI system recommends more diagnostic tests, referrals or prescriptions than the case needs. It often looks like thoroughness, such as a routine lab panel on every visit. Accuracy benchmarks rarely penalize it, so it has to be measured directly against what clinicians would order and whether each addition is justified.
How much more do AI care plans cost than physician plans?
In a NEJM AI study published October 15, 2026, 24 AI systems reviewed 200 outpatient and emergency notes. Their plans averaged $314 per visit in tests and referrals, against $97 for the treating physicians, about 3.2 times more. System averages ranged from $209 to $599, as reported by TechTarget.
Does prompting a model to be cost-conscious solve overtesting?
Not by itself. In the NEJM AI study, cost-conscious prompting lowered cost ratios across 19 systems, but 17 of them showed lower agreement with physician diagnoses, per 2 Minute Medicine. Any change meant to reduce orders should be retested on the same physician-graded set for both cost and accuracy.
Is the physician's plan the right benchmark for an AI care plan?
It is a useful comparator, not a gold standard. Physicians can undertest, and some AI additions are appropriate. In the NEJM AI study, nearly a fifth of AI plans matched guidelines and caught orders the physician missed. That is why every added order should be judged by specialty-matched physicians, not just counted.
Who should grade an overtesting evaluation?
Practicing physicians matched to the specialty and setting of each case. Two or more should label every added test, referral and prescription as necessary, reasonable or unnecessary, with a third resolving disagreements. Stewardship or utilization leaders can supply the low-value order lists and the fee schedule used for pricing.
How often should we rerun an overtesting evaluation?
After every model, prompt, retrieval or order-set change, and on a regular schedule after go-live. Model behavior shifts between releases, and small changes can move order rates. Keep the test set private, refresh part of it each quarter, and pair it with production monitoring of order and referral rates.