The Same Patient, Two Prompts, Two Different Answers

Sent to subscribers on 28 September 2026. Details may have changed since.

Training, research, and the specialty itself. Three things worth knowing this week.

Residency and Fellowship · The residency rulebook is being rewritten, and comment closes October 22
Research · A heart team's cases, handed to two AI models, twice
Cardiothoracic Surgery News · TAVR without the surgeon in the room, DCD hearts at scale, and the appendage again

Part One · Residency and Fellowship

Fewer Rules, More Discretion

Programs opened applications last week. While you were refreshing your inbox, the organization that decides what your life looks like after you match published something worth your attention. On September 8, the ACGME posted a major revision of the Common Program Requirements, the document every accredited residency program in every specialty must follow. It is open for public comment until October 22. If the ACGME's Committee on Requirements and Board adopt it at their June 2027 meetings, it takes effect July 1, 2028.

That date matters. If you match in March 2027, you will do your intern year under the current rules and switch as a PGY-2. If you are applying in the 2028 cycle, this is your rulebook from the first day. Here is what is in it, taken from the ACGME's own rationale and impact statement.

The scale of it

43% fewer requirements than the current document
July 1, 2028 proposed effective date

The dates that matter

Oct 22 public comment closes
June 2027 final draft goes to the ACGME board

What changes for the resident, specifically

Who runs your program. Program directors no longer need Review Committee approval, only the institution's GME committee, and the three year minimum of documented educational experience is gone, replaced by a mentoring plan for anyone with less. A Review Committee may permit co-program directors where that is needed for residents to be eligible for every board certification option, though the ACGME expects nearly all committees not to. Leadership becomes a local decision.

How big and how staffed. Review Committees lose the power to set a minimum resident complement and can no longer require more than one core faculty member per five residents. Both are framed as removing barriers to new programs in rural and underserved areas. Expect more small programs, and expect them to be leaner.

Pregnancy and lactation. Every program must have a written policy for pregnant residents covering schedule modification and occupational exposure, and must provide protected time and private refrigerated facilities for lactation, with no requirement to make the time up. The pregnancy policy is new at the common requirement level; the lactation requirement is expanded.

Board eligibility, disclosed at interview. Programs already have to tell applicants offered an interview about their eligibility for the specialty board examinations. The revision widens this to eligibility and options for the relevant certifying boards, and adds residents when they start the program. Ask this season. A program that cannot answer cleanly is telling you something.

Artificial intelligence, named. The professionalism competency now explicitly includes ethical use of technology, and the text names artificial intelligence. So does the systems-based practice competency, under the use of existing and emerging technologies.

The hours. The 80 hour week, averaged over up to four weeks, remains in the proposed text at requirement 6.29, now averaged within a single rotation. I am not going to characterize the rest of Section 6 beyond what I confirmed line by line. Read it yourself; it is the part everyone will go to first.

What this is, in one sentence: less prescription from the accreditor, more discretion for your program and your institution. That cuts both ways. A program with a strong culture will use the room well. A program that only did the right thing because a requirement forced it now has less forcing. The rulebook was never what protected residents in good programs. It was what protected residents in the others.

Fewer rules are good news for the programs you would want to train in and mixed news for the ones you would not. Which kind you end up in is being decided this interview season.

What to do with this

Two things. First, read Section 6 of the proposed text and submit a comment before October 22 through the ACGME review and comment page. Second, turn the proposal into interview questions. Ask how the program supports pregnant residents. Ask what the core faculty to resident ratio is. Ask how they would use more discretion if they had it. The answers tell you what the culture will do once the rulebook stops forcing it.

Part Two · Research

The Same Patient, Two Prompts, Two Different Answers

In Melbourne, a multidisciplinary heart team meets every week to decide what to do with patients who have complex coronary disease. Five hundred forty six of the patients it discussed between 2019 and 2024, mean age 66, three quarters with triple vessel disease, 12 percent with severe left main disease, went into a study published online in JSCAI. The team had recommended bypass in 67 percent, medical therapy in 19 percent, PCI in 13 percent, and 91 percent of patients received what the team recommended.

The investigators then handed every case to two large language models, ChatGPT-4o and Gemini 2.0, and asked what they would recommend. They asked twice. Once with the free text clinical narrative the heart team had actually read. Once with a structured pro forma of 30 clinical and angiographic variables, the kind of tidy table a registry or an EHR extraction would produce.

Given the clinical narrative

75% agreement with the heart team

Given the 30 variable table

35% agreement with the heart team

Same patients. Same models. Same question. The only thing that changed was the shape of the information, and agreement with the surgeons and cardiologists in the room fell by more than half. Adding the European and American revascularization guidelines to the prompt did not help. The senior author's summary, to TCTMD: "The output that we got from AI was entirely dependent on the context that we gave it."

There is a second finding, and it is the one I would carry into a meeting. When ChatGPT-4o disagreed with the heart team, those patients had more major adverse cardiovascular events over a median of 989 days, odds ratio 1.67, interval 1.03 to 2.72. Before you conclude the model saw something the team missed: after adjustment for clinical complexity and the treatment actually given, the association disappeared, and it was never present for Gemini. The disagreement was not insight. It was a marker. The cases where a model and a heart team part ways are the cases that were hard in the first place.

A model does not know what it was not told. It will answer anyway, and it will sound exactly as confident either way.

Why does the table lose to the narrative? Because the narrative carries what the table cannot. The frailty the surgeon noticed walking the patient to the exam table. The porcelain aorta mentioned in passing on the CT report. The patient who said she would not accept a sternotomy. The diffuse distal disease that made grafting unattractive. Thirty variables is a lot of variables and it is still not the case. Every surgeon knows this. What the study adds is that the machine, in effect, knows it too, in the sense that its performance collapses without that information. It just does not tell you that it collapsed.

Three rules for asking a model a clinical question

1. Give it the story, not the summary. Write the case the way you would tell it to a colleague at the scrub sink, including what worried you. In this study the structured fields alone cut agreement with the experts by more than half.

2. Treat disagreement as a flag, not a verdict. When the model and the team disagree, you are most likely looking at a high risk patient, not a wrong team. That is the moment for a second human opinion.

3. Never let the tool see less than the human saw. If a variable mattered to your decision and it is not in the prompt, the answer you get back is an answer about a different patient.

The usual caveats apply and they are real. Single center, retrospective, the heart team as the reference standard rather than the truth, and two models that will be superseded within months. None of that weakens the operational lesson, because the effect is too large to be an artifact of any of them. The same is true of the paper's own conclusion, which asks for "format aware" integration and validation before clinical deployment. Format aware is a polite phrase for: what you feed it decides what you get.

What to do with this

Run your next question twice. Whatever AI tool you use for a clinical or research question, ask it once as bullet points and once as a paragraph with the context you would give a colleague, then compare. If the answers differ, the difference is the information you left out, and it was probably the information that mattered. This is also how AskDimitri is meant to be used: the more of the actual situation you give it, the more of the actual answer you get back.

Part Three · Cardiothoracic Surgery News

What Moved in September

Three items with direct consequence for how cardiac surgery is organized, not just how it is performed.

CMS · Final national coverage determination for TAVR, CAG-00430R2 · September 10, 2026

Medicare no longer requires a cardiac surgeon in the TAVR room, or a hospital SAVR volume to host one

After a nine month review opened in December 2025 and 257 public comments, CMS finalized the first rewrite of the TAVR coverage rules since 2019. The changes: TAVR may be performed by a single operator, cardiologist or surgeon, where the 2019 rule required both to jointly participate in the procedure. Hospital volume thresholds are gone; the 2019 rule required at least 50 aortic valve replacements a year including 20 TAVRs, or 100 over two years including 40, plus 300 PCIs. Operator volume survives at 20 transcatheter valve procedures a year, 15 of them TAVR, or 40 over two years, 30 of them TAVR. Symptomatic severe aortic stenosis is now covered without coverage with evidence development. Asymptomatic severe aortic stenosis is covered only under CED, and one of the evidence questions compares TAVR, SAVR, and surveillance until symptoms. Mandatory registry participation is replaced by a "continuous quality improvement process." What survived: a heart team whose evaluation must include input from both a cardiac surgeon and an interventional cardiologist, though it may be asynchronous, and on-site cardiac surgery. STS said publicly the policy "does not incorporate all of STS' recommendations regarding cardiac surgeon involvement." The practical read: the surgeon's role in the TAVR pathway moves from the table to the record. Whether that is fine depends entirely on how seriously your institution takes the record.

Journal of Cardiac Failure · September 2026 · Thomas Jefferson University, Philadelphia

DCD heart transplantation spread to more centers, and the outcomes did not move

1,489 adult recipients of hearts donated after circulatory death in the UNOS registry, January 2019 to June 2024: 303 transplants in the early era, 2019 to 2021, and 1,186 in 2022 to 2024, by which time the technique had spread across many more transplant centers. No significant difference in one year mortality, retransplantation, length of stay, or dialysis before discharge, before or after inverse probability weighting, and none in the high risk subgroups, pulmonary hypertension, preoperative mechanical support, impaired functional status, even though late era recipients had broader risk profiles. The accompanying editorial is titled "The DCD Juggernaut." For the surgeon, the finding is unusual: a technique that a few centers pioneered in 2019 has been reproduced nationally without an early outcome penalty, which almost never happens with a new operation. The authors conclude that it was adopted with enough procedural maturity from its early phase to support continued expansion. Remember that the next time someone says a technique cannot be taught at scale.

JSCAI · Bayesian meta-analysis, six randomized trials · September 2026

Percutaneous appendage occlusion: a 94 percent probability of more ischemic strokes than anticoagulation

7,004 patients with nonvalvular atrial fibrillation in six randomized trials of percutaneous occlusion versus oral anticoagulation, more than 18,000 patient years of follow up. Ischemic stroke incidence rate ratio 1.34, a 94 percent posterior probability that the device arm had more. Hemorrhagic stroke went the other way, rate ratio 0.55, and that advantage shrank to 0.80 when only trials with a DOAC comparator were pooled. Any stroke, 1.09. The authors call the absolute excess small and say it has to be weighed against fewer hemorrhagic strokes. For a surgical reader the value is the contrast. LAAOS III cut ischemic stroke or systemic embolism by a third with surgical occlusion added to usual care, most patients staying on anticoagulation. LAACS-2, covered in Issue 17, did not support routine closure during cardiac surgery in all patients, with or without atrial fibrillation, though those at high stroke risk may still benefit. Now a pooled device analysis leans toward more ischemic events than anticoagulation alone. The appendage question is not one question. It is at least three, and they keep being answered separately.

Next Monday's letter is the last before the 40th EACTS Annual Meeting opens in Barcelona on October 7. Until then, whether you are deciding what to ask a program this interview season, how much to trust an answer a model gave you, or what a coverage rule changes in your own operating room, that is the kind of thinking AskDimitri is built for.

Work through it with AskDimitri →

Editor's note: in October 2026 the figures in this letter were checked against the sources listed below, and the text was corrected where the letter as sent differed from them. Figures that could not be verified were removed: the transplant center counts and the era-by-era mortality, length of stay and dialysis figures in the DCD item, the graft failure and survival figures that could not be read in the paper itself, and the list of trials, devices, follow up and the interval in the appendage item.

Until next week,

Sources

A new letter goes out every week, to subscribers first. Join to get the next one.

Fewer Rules, More Discretion

The Same Patient, Two Prompts, Two Different Answers

What Moved in September

Medicare no longer requires a cardiac surgeon in the TAVR room, or a hospital SAVR volume to host one

DCD heart transplantation spread to more centers, and the outcomes did not move

Percutaneous appendage occlusion: a 94 percent probability of more ischemic strokes than anticoagulation

Sources

Get the letters, and access to the assistant