Study 1 — the randomized trial (breast oncology, n = 124)

Published on March 12, 2026, in Frontiers in Medicine. One hundred twenty-four breast oncology residents from three urban tertiary teaching hospitals in the same metropolitan region, all nationally accredited centers for standardized residency training. Individual 1:1 randomization by computer-generated sequence prepared by an independent statistician, with allocation concealment via sequentially numbered, opaque, sealed envelopes. Sixty-two in the AIEIT arm, sixty-two in conventional training.

Assessors for the subjective outcomes — clinical reasoning, case presentation, teamwork — were blinded to allocation and were not involved in delivering the intervention. Participants and instructors were not blinded, which is unavoidable in an educational intervention. Prior sample size calculation: expected effect d ≈ 0.6, two-sided alpha 0.05, 80% power, requiring 58 per group. The intervention spanned a three-month full-time clinical rotation.

The intervention had four components: a dynamic knowledge graph using natural language processing over NCCN guidelines, updated monthly; a mixed-reality multidisciplinary team platform with roughly 20 standardized scenarios; a virtual patient–AI mentor system with roughly 25 cases; and a learning analytics dashboard. Declared primary outcome: knowledge acquisition and clinical reasoning performance.

Study 2 — the transesophageal echocardiography simulator (anesthesiology, n = 60)

Published on May 25, 2026, in Frontiers in Surgery. Single-center randomized study at Sichuan Provincial People’s Hospital, between December 2024 and December 2025. Sixty anesthesiology residents, thirty per arm. Both groups received the same instructor-led TEE curriculum — the AI arm additionally trained on a simulator with 3D anatomy visualization, interactive image-reading exercises, and virtual probe manipulation with real-time feedback.

This is the cleanest of the three designs, and it is worth understanding why: because the curricular base is identical, the observed difference is attributable to the addition of the simulator, not to “having technology versus having nothing”.

Study 3 — the obstetrics and gynecology study (n = 28)

Published in 2026 in BMC Medical Education. And here is the difference that changes everything: it is not a randomized trial. It is a prospective interventional study with a pre-post design, with no control group. Twenty-eight residents from three consecutive cohorts — 2021 (n = 10), 2022 (n = 10), 2023 (n = 8) — at a single hospital, the First Affiliated Hospital of Fujian Medical University. The total period was 36 months, and the principal comparison covers 12 months of platform-based training.

Same family of intervention — case simulation platform, adaptive-algorithm personalized pathway, competency-based assessment system. A full rung lower on the evidence hierarchy.

All four domains in the OB/GYN study report Cohen’s d between 1.58 and 2.56. Overall competency: 2.30. Clinical skills: 2.56. Clinical reasoning: 2.36. Professional knowledge: 2.02.

These are enormous effects. And the design that produced them is pre-post with no control group, measuring 12 months of residency.

Ask the obvious question: what else happened to those 28 residents over twelve months? They did their residency. They saw patients, worked shifts, were supervised, read, made mistakes and corrected them. A first-year resident improves substantially from month 1 to month 12 with or without an AI platform. With no control group, there is no way whatsoever to separate the effect of the intervention from the effect of simply having completed another year of residency.

Add to that two well-described mechanisms that systematically inflate pre-post designs: the testing effect — anyone who has taken the assessment once performs better the second time, even without learning anything — and the Hawthorne effect, the performance gain that follows from knowing one is being observed within an innovation project. The authors of the oncology randomized trial explicitly state that they cannot exclude a Hawthorne effect in their own trial, which does have a control group. In the uncontrolled study, the hypothesis cannot even be raised.

None of this means the intervention does not work. It means the OB/GYN study cannot measure how much it works — and that d = 2.30 should be read as the combined effect of intervention, maturation, testing and observation, not as the effect of the platform.

Here the reading turns practical, and the conclusion is counterintuitive.

Look at what is expensive in the randomized trial’s intervention: mixed reality, twenty multidisciplinary team scenarios, twenty-five virtual patients, natural language processing over guidelines refreshed monthly. That is a months-long engineering project with a budget.

Now look at what the authors themselves identify as the mechanism of the effect, in their discussion: personalization of the learning pathway, immediate feedback in a risk-free environment, and the shift from episodic summative assessment to continuous monitoring with targeted remediation. Translated: dense, individualized, high-frequency feedback.

That is exactly what a good attending does. What the platform solves is not the nature of the feedback — it is its scale. And what AI made radically cheaper was not teaching: it was the instrumentation of teaching. Measuring competency before and after, with a validated rubric and paired analysis, costs a spreadsheet and thirty lines of code today.

Apply this to what happens in your own cohorts. Two AIMED cohorts for OB/GYN residents, in May and June, mixed across first to third year — had they been instrumented with a rubric applied on entry and exit, they would have generated the same level of evidence as the Fujian study, with two advantages that study does not have: two independent cohorts rather than three cohorts from the same service, and — more importantly — training year as a clean stratification variable rather than a confounder.

The difference between data and publication, in this case, is not methodological rigor. It is instrumentation — and instrumentation is a decision made before the class starts, not after.

CFM Resolution No. 2,454/2026, published in Brazil’s Official Gazette on February 27, 2026, regulates the use of artificial intelligence in medicine. It classifies systems by risk level — low, medium, high or unacceptable — guarantees that the final word on diagnosis, therapy and prognosis always belongs to the physician, prohibits delegating the communication of diagnoses and prognoses to AI, and bars institutions from penalizing physicians who do not follow the machine’s recommendation.

Note what it regulates: clinical decision-making. And note what it does not, by construction, reach: AI applied to medical training. A virtual patient simulator training a resident issues no diagnosis about a real patient. A learning analytics dashboard tracking competency influences no therapeutic conduct. None of the three studies in this article describes a system that CFM 2.454 would classify.

At ANVISA, the Brazilian health regulator, the reasoning arrives at the same place by another route. Regularization of software as a medical device depends on the software aiding diagnosis, monitoring a clinical condition, or influencing therapeutic decisions. An educational platform operating on simulated cases does not meet that trigger.

The correct reading of this is not “it is unregulated, do as you please”. It is the opposite: this is simultaneously the fastest-adopting and least-regulated area of medical AI. When a residency program adopts a platform that decides what each resident studies and how each resident is assessed, no CFM resolution and no ANVISA registration demands prior validation. The only barrier is the quality of evidence the program itself requires before adopting — and that barrier is institutional, voluntary, and today all but nonexistent.

Cardiology: the most favorable case of all, and almost no one has noticed

Cardiology has a property that oncology and gynecology do not: three core competencies with an objective, digitizable and scalable gold standard.

Electrocardiography. It is the only high-volume clinical competency in which the assessment item is a file, the answer key is consensual, and the case bank can hold tens of thousands of examples with known outcomes. Nothing equivalent exists in gynecology. An adaptive system serving the next tracing calibrated by the resident’s previous error is, technically, the simplest application in this entire article — and the one with the highest repetition density.

Echocardiography. Here direct, favorable evidence already exists, albeit outside cardiology: the transesophageal echocardiography trial with sixty anesthesiology residents showed gains in key-view acquisition (88.2% versus 76.5%) and reduced procedure time (27.1 versus 32.4 minutes), with an identical curricular base in both arms. Transthoracic echocardiography in cardiology residency has the same profile: view acquisition is objectively measurable, and OSATS is a validated, transferable instrument.

Hemodynamics and electrophysiology. These are the domains in which real exposure is scarcest and most dependent on institutional volume — precisely the condition in which a simulator delivers the greatest marginal value. And it is where the competency reference has just been updated: the 2026 ACC/AHA/HRS Advanced Training Statement on Clinical Cardiac Electrophysiology revised training requirements and designated competencies with explicit alignment to cutting-edge technologies.

The detail that inverts the order of priority

Intuition says “start with the ECG, it is the easiest”. Reading the three studies suggests the opposite.

Note which domain gained most in the OB/GYN study: clinical reasoning, which started from the lowest baseline score (68.4) and rose the most (16.5 points). The authors record that higher-order cognitive skills — risk assessment accuracy and comprehensive differential diagnosis formulation — had the lowest initial scores and the greatest improvement margins, 27.4% and 24.8%. The same pattern appears in the oncology trial, whose most pronounced gains were in specialized oncology reasoning and multidisciplinary collaboration, not in isolated technical tasks.

In other words: the effect is largest where the competency is most complex and least trainable by repetition, not where it is easiest to digitize. In cardiology, that shifts the bet from the ECG in isolation toward decision-making under uncertainty — the syncope case discussion, the device indication, preoperative risk stratification, the management of intermediate-risk chest pain. That is the territory where variability between cardiologists is greatest and the rubric is hardest to write. It is also where the gain, on the available data, would be largest.

This is an uncomfortable result for curriculum planners, because it inverts the order of least effort. But it is what the two studies with domain-level data show.