Nov 19, 2026  ·  4 min read

AI drafts of heart ultrasound reports match human ones, a trial says

A receipt spike on an oak base, holding blank paper slips and one ink-blue slip

Cardiologists changed AI drafts about as often as human drafts and checked them faster, but the result is preliminary.

An AI that drafts heart ultrasound reports

On Nov. 9, researchers at Kaiser Permanente Northern California presented trial results at the American Heart Association's meeting in Chicago. They tested an AI that writes the first draft of a heart ultrasound report, known as an echocardiogram. A cardiologist, a heart doctor, then checks the draft and signs the final version. The researchers say the AI drafts were about as good as human drafts, and checking them took less time. That matters because heart ultrasound results often wait in a cardiologist's queue.

Three ways to write the first draft

Sonographers, the technicians who run the ultrasound probe, can draft a preliminary report for the cardiologist. For this trial, the team took 1,200 heart ultrasounds already made in routine care. Each one went through three workflows. In one, the AI wrote the first draft. In the other two, a sonographer or a cardiologist other than the signer did. The study had no patient contact and no effect on anyone's care.

A cardiologist then finalized each draft without knowing who or what had written it. The AI's drafts were made to look human, so the format gave nothing away.

“I believe the result answers its narrow question, but I want the full paper and patients' waiting times before I read more into it.”

Scoring a draft by how much gets rewritten

The trial counted how often a cardiologist made a substantial change to a draft. Say a draft gives the share of blood the heart pumps out per beat. Moving it by 10 points or more counts. A leaking valve getting a new severity grade also counts, even a one-step shift. One such change is enough to flag a whole draft.

Cardiologists made such changes to 50.6% of the AI drafts and 49.7% of the human drafts. So about half of drafts changed either way, a gap of under a point. The team had set a margin of 5 points in advance: the AI could trail by that much and still count as a match. It stayed inside.

The investigators do not claim the AI beats people, only that it is not meaningfully worse. One commentator adds a catch. The main measure shows how often cardiologists changed a draft, not how often a draft or final report was right. Another commentator says the high rate for both kinds of draft suggests the final check does real work.

Minutes back for cardiologists

Cardiologists spent an average of 4.6 minutes finalizing an AI draft. Without AI help, writing a report took 8.5 minutes. That is a time saving of about 46%, or about 4 minutes a study.

Keane K. Lee, a cardiologist and co-senior author, says a few minutes saved per study add up over a day, and for patients that could mean faster results. The trial authors say the saved time could also let each cardiologist read more studies, easing a shortage of specialists.

The study timed only the cardiologists' part. It did not measure how soon patients got their results, or how they fared later. I read faster results and higher volume as different gains, and a shorter read proves neither.

The fine print, and who signs off

The numbers come from a conference abstract. It has not had peer review, a check by independent experts, and the full paper is not out. The data come from one health system. The cardiologists also knew they were in a study, which may have changed how they worked.

Some authors have ties to the AI they tested. The model, EchoPrime, belongs to Cedars-Sinai Medical Center, which gave InVision Medical Technology Corporation the exclusive right to sell it. Authors David Ouyang, Bryan He and Milos Vukadinovic report rights to the model, and He and Vukadinovic also report a financial stake in InVision.

Hani Najm, a vice president of the American College of Cardiology, says the study showed AI readings were accurate enough. He adds that the medical community needs to set guardrails on AI use.

I like the split of duties here. The AI wrote page one, and a person owned the last page. Lee says the cardiologist stays responsible for certifying that the final report is accurate, even when AI helps. I suspect that split is easier to keep in a trial than on a busy ward. There, a rushed doctor may skim and sign. That is where Najm's guardrails would matter. One commentator says patients can ask who will review their results and whether software helped prepare the report.

Nov 19, 2026

  1. Nov 19Google rolls out Genie 4 to eligible Ultra subscribers
  2. Nov 19FCC sets December 17 compliance date for SIM swap fraud rules
  3. Nov 19GAO says four major AI chatbots produced viable attack plans
  4. Nov 18Traitomic raises €20 million to expand its genetic screening business
  5. Nov 18GitHub rebuilds Copilot's agent layer so sessions move between devices

This is our sketchbook.

Two stoneware cups turned towards each other, as if in conversation

Got a good project?
Let’s work on it together.

Let’s talk →

info@ndot.io · New Jersey · Budapest · Biel