Forwarded this by a colleague? Subscribe free and get it every week.
THIS WEEK
Epic's rebuilt deterioration score is FDA-cleared. Clearance does not show that it saves lives. This week: what Epic's testing tells us, what to ask before rollout, and a practical way to check a clinical trial with AI.
In this issue:
Epic's Deterioration Index v2 is FDA-cleared
A Baltimore hospital switches on an FDA-authorized sepsis score
CHOP builds pediatric heart models in seconds
In the literature: a MAP alarm versus HPI, what 952,000 clinicians ask OpenEvidence, and what LLM-recommended care costs
Spotlight: five AI terms worth knowing
How-to: read a clinical trial like a skeptic, with AI's help and a free field guide
LATEST NEWS
Epic's deterioration score is now an FDA-cleared device
The Deterioration Index on many inpatient lists has run in hospitals since 2017 without FDA review. On September 18, the FDA cleared a rebuilt Version 2 (510(k) K260008).
What's new. The old score predicted escalation of care within 12 hours or death within 38 from a smaller set of vitals, labs and nursing assessments (Byrd et al.). Version 2 uses 195 pieces of chart data, including medications, diagnoses and trends. It predicts a rapid response call, ICU transfer or code within 24 hours, or death within 72, for adult inpatients outside the OR, PACU and labor and delivery.
How well it works. Epic tested it at three hospitals on about 10.7 million scores; 3.8% were followed by one of those events. A cutoff of 70 caught about half of them, and one alert in five was followed by an event. A cutoff of 90 cut alerts by two-thirds but caught only about a quarter. More alerts or more misses: every hospital will have to choose.
A caution on equity. Pooled across hospitals, the score caught 26% of events in Black patients versus 51% in White patients. Within each hospital the gap mostly closed; most Black patients were at the site where the score did worse for everyone. The same score can perform very differently from one hospital to the next.
Why it matters: FDA clearance means the score matched a similar cleared tool on past data. It does not show that it saves lives. Before your hospital switches it on, ask what score triggers a page, who answers it, how many alerts a nurse will get per shift, and whether it was first run silently on your own patients, with no alerts, to see how it performs here.
A Baltimore hospital switches on an FDA-authorized sepsis score
Sinai Hospital in Baltimore has started using Prenosis's Sepsis ImmunoScore, LifeBridge Health announced. It was the first AI sepsis tool authorized by the FDA (De Novo, April 2024), labeled for ED and admitted patients with suspected sepsis and a blood culture already ordered. In external validation of 698 such patients (Bhargava et al., NEJM AI), 37% of the high-risk group and 70% of the very-high-risk group developed sepsis within 24 hours, versus 3% of the low-risk group. Prenosis helped fund the study, and several authors work for the company.
Why it matters: The score fires only after someone already suspects sepsis. It stratifies; it does not find the patient nobody is worried about. Ask whether a high-risk result changes antibiotic timing or level of care, or just adds a label.
CHOP builds pediatric heart models in seconds
Children's Hospital of Philadelphia now builds patient-specific 3D heart models from CT, MRI and 3D echo in seconds rather than hours, Isha Salian reports on NVIDIA's blog. Matthew Jolley, a pediatric cardiologist and cardiac anesthesiologist, led the work in his group's open-source SlicerHeart toolkit. CHOP now models complex VSDs before surgery routinely. The source is vendor content; its only number is speed, with no published accuracy or outcome data.
A 3D heart model built in SlicerHeart, the open-source toolkit from Jolley's group, used to plan a surgical repair. Lasso A, Jolley MA, et al. Front Cardiovasc Med. 2022. CC BY 4.0.
Why it matters: Faster models make virtual rehearsal practical for more congenital cases, anesthetic plan included. Ask how often a model has changed the operation.
IN THE LITERATURE
Anesthesiology — Florax et al.
Design: Single-center, blinded noninferiority RCT in the Netherlands; 143 adults having elective noncardiac surgery with arterial lines.
Finding: A MAP alarm below 72 was noninferior to HPI above 85 for hypotension burden (median 3.75 vs. 4.00 mmHg·min under MAP 65). Organ injury and 30-day mortality did not differ.
Limitation: One center; not powered for organ outcomes.
Vendor: HPI is Edwards Lifesciences' algorithm. Funding is not stated in the abstract.
Bottom line: With Issue 4's Wu trial, neither small RCT found an advantage for HPI over a MAP trigger in the low 70s. Worth discussing before your department buys HPI.
JAMA — Patel et al.
Design: Cross-sectional study of OpenEvidence query logs from 952,008 NPI-verified U.S. clinicians, January 2024 to June 2026.
Finding: 271 million queries. Treatment and management accounted for 33.1% of physician queries; diagnosis, 13.9%. Anesthesiology's use relative to its workforce share was 0.37.
Limitation: It shows what clinicians asked, not whether the answers were correct or improved care.
Vendor: Three authors work for OpenEvidence, which supplied the data.
Bottom line: Treatment questions dominate. Whether the answers to them are right is the next thing to study.
NEJM AI — Aran et al.
Design: 24 AI systems wrote plans for 200 deidentified outpatient and ED notes from a public collection; orders were priced at 2026 Medicare rates.
Finding: Plans from general-purpose systems cost a mean $314 per visit vs. $97 for physicians' documented plans, 3.2 times as much. Where physicians ordered no tests, every system suggested some. A "parsimonious but safe" prompt cut the ratio to 1.4.
Limitation: Sample notes, no patient outcomes; cost is not value.
Funding and conflicts: Not stated in the public abstract; disclosure forms are posted with the article.
Bottom line: Unprompted, LLMs tend to order more. If you use one for a plan, tell it to order only what's needed, and decide what is indicated yourself.
SPOTLIGHT
Five AI terms, explained for clinicians
These words come up in vendor demos and committee meetings. Here is what each means at work.
Large language model (LLM). The engine behind ChatGPT, Claude, Gemini and OpenEvidence. It writes by predicting what text comes next from patterns learned in training, not by looking facts up. It can sound knowledgeable and still be wrong. Example: ask for the dantrolene dose in malignant hyperthermia and you get a fluent answer because that pattern is common in its training text, not because it checked a reference.
Tokens. The words or word fragments a model processes, about three-quarters of a word each. They set how much a model can handle and what it costs. Example: "anesthesiologist" may be split into several tokens, and a 10,000-word chart summary runs roughly 13,000.
Context window. What the model can keep in view at once, measured in tokens. In long chats, details get dropped or missed. Restate key facts or start fresh. Example: paste in a long ICU course, mention a penicillin allergy at the start, and an hour of questions later the model may suggest a drug as if it never saw it.
Hallucination. A confident, false answer. Check numbers, doses and citations against their sources. Example: ask for references on sugammadex and bradycardia and you may get a real journal, plausible authors, and a DOI that leads nowhere.
Retrieval-augmented generation (RAG). The system searches a set of sources first, then answers from what it found, with citations. That reduces hallucination but doesn't end it: the search can pull the wrong passage, and the model can misread the right one. Example: a hospital chatbot that searches your policy library before answering a question about transfusion thresholds. If it retrieves last year's version of the policy, it will cite it confidently.
The takeaway: Each term is a question for the next AI tool pitched to your department. Which model is it built on? How much of the chart fits in its context window? What does it retrieve from? How often does it make things up? Then ask to see the evidence behind the answers.
HOW-TO
Before you try this: check your organization's approved-tool list, and never put patient information into a tool it hasn't approved.
Read a clinical trial like a skeptic, with AI's help
A trial lands in your inbox and the abstract says the drug worked. Use AI to extract information, then check it yourself. In an exploratory comparison by Gartlehner and colleagues in Annals (63 studies, 9,341 data elements), AI-assisted extraction with human checking was 91% accurate versus 89% for a human-only workflow. Separate studies found limits elsewhere. None of the four models tested by Nyrhi and colleagues in eBioMedicine could rate risk of bias on its own across 200 trials. Models tested by Yun and colleagues recognized spin when asked yet still carried it into summaries. Use the AI as extractor and devil's advocate. You stay the judge.
As of October 2026, Claude, ChatGPT, and Gemini all read uploaded PDFs. Published papers rarely contain patient identifiers, but some include consented case details or images, so check the files and follow your institution's policy.
Write a first impression. After a quick read on your own, jot one tentative sentence about what the trial shows and what you're unsure of. Revisit it at the end. Agreement deserves scrutiny too.
Gather the documents. The full paper, supplement, protocol and statistical analysis plan if available, and the trial registration. On ClinicalTrials.gov, use Record History to see the original registration, not just the current one.
Extract with receipts, harms included. Paste the prompt below.
You are helping a clinician appraise a randomized trial. First, list the files you can read and any pages or tables you cannot. Then, using only these documents, extract the items below. For each, quote the text and give the document, page, and table or figure. If an item is absent, write NOT FOUND IN SUPPLIED DOCUMENTS. If a needed document is missing or unreadable, say so. Do not infer or estimate.
1. Population: key inclusion and exclusion criteria
2. Intervention and comparator: dose, timing, co-interventions
3. Hypothesis: superiority, noninferiority (with margin), or equivalence
4. Primary outcome: definition, time point, original vs. current registration
5. Numbers randomized and analyzed per arm; denominator for each outcome
6. Primary result per arm, effect estimate, and confidence interval at the reported level
7. Harms: serious adverse events, key safety outcomes, and discontinuations for adverse events, with denominators and follow-up
8. Composite outcomes broken into components; patient-important vs. surrogate outcomes
9. Randomization, allocation concealment, who was blinded
10. Loss to follow-up; ITT, modified ITT, or per-protocol analysis
11. Funding and author conflictsCheck the claim against the hypothesis. Ask whether the trial tested superiority, noninferiority or equivalence. For noninferiority, have it state the margin and whether the confidence bound clears it; a positive noninferiority conclusion is not spin. Then compare the conclusions with the prespecified primary analysis and the harms.
Run a bias check informed by RoB 2. Ask for the relevant passages, missing information and provisional concerns in each domain. Verify the evidence behind any pivotal rating, low as well as high.
Calculate only what fits. Have it name the outcome type and the right denominator first. Risk differences and NNT suit binary outcomes over a fixed period; a hazard ratio is not a risk ratio. Report an NNT only with its interval and time horizon, and only when the risk difference excludes zero.
Verify, then summarize. Check the primary result, denominators, confidence interval and key harms against the paper yourself. Then ask for five bullets: the question, benefit with its uncertainty, harms, the main concern, and what is still unresolved. A second model, run independently, can catch extraction errors. But in Khan and colleagues' study, even answers two models agreed on were wrong 6% of the time.
When to stop. Don't change practice on an AI appraisal alone. If its read and the paper disagree, the paper wins. Treat this as a 20-minute first pass, not a full appraisal.
All six prompts, including harms and a noninferiority check, are in the guide. Get the free field guide: Read a Clinical Trial Like a Skeptic.
I have no financial relationship with OpenAI or Google. Anthropic makes Claude, one of the AI tools that help draft this newsletter; I pay for it like any subscriber and have no other relationship with the company.
THAT'S IT FOR THIS WEEK
How was this issue?
Send us a tool or a how-to. Using something in clinic that other clinicians should know about? Submit it here and it may show up in a future issue. We ask before we publish anything and credit you by name only if you want us to.
Know one colleague who’d use this? Forward this email, or send them the link below.
Loading Dose AI is for educational purposes only and is not medical advice. Nothing here should guide the care of a specific patient, and reading it does not create a physician-patient relationship. AI tools help gather and draft each issue; I read every source and edit every word. Opinions are mine and not my employer's. Full disclaimer