The first randomized trial of AI in cardiology.
Most medical AI reaches the market on retrospective evidence alone. InVision's LVEF model was tested the way a drug is tested: prospectively, blinded, and individually randomized against the human standard of care — then published in Nature.
EchoNet-RCT — Nature, 2023
A blinded, individually randomized trial comparing AI initial assessment of left ventricular ejection fraction against sonographer initial assessment, with the reviewing cardiologist blinded to which arm produced the draft.
| Item | Value |
|---|---|
| Primary result | Difference of −10.4 percentage points (95% CI −13.2 to −7.7) in the proportion of studies substantially changed by the reviewing cardiologist |
| Significance | P < 0.001 |
| Analyzable studies | 3,495 transthoracic echocardiograms |
| Design | Blinded, individually randomized, AI versus human initial assessment |
| Registration | NCT05140642 |
| Publication | Nature, 2023 (He et al.) |
"Substantially changed" is the trial's own endpoint: a change by the reviewing cardiologist large enough to matter clinically. Fewer substantial changes means the draft the cardiologist received was closer to the read they signed.
Why the design matters
Retrospective studies show that a model can reproduce a label on data it did not train on. That is necessary, and it is not the same question a health system is asking. The clinical question is whether putting the model into the workflow changes what the cardiologist does — and only a prospective trial with a concurrent control arm can answer it.
Three features of this trial are worth checking against any comparison you are given. It was prospective, so the outcome was not selected after the fact. It was individually randomized rather than split by time period or by site, which removes the confounding a before-and-after comparison carries. And it was blinded — the reviewing cardiologist did not know whether the draft in front of them came from the AI or from a sonographer, so the comparison measures the draft rather than the reader's expectations of it.
Validation studies
Trial evidence is one tier of the ladder. Below it sits prospective external validation, then retrospective multi-site, then single-site. Here is where each product's evidence actually sits.
| Product | Strongest evidence | Design |
|---|---|---|
| Precision LVEFFDA cleared | EchoNet-RCT, Nature 2023 | Blinded randomized trial. The strongest design in this category. |
| Precision Cardiac AmyloidFDA cleared | JACC: Advances 2025 | Retrospective external validation across 5 sites in two countries, plus an independent head-to-head run by investigators outside InVision. Full evidence → |
| Precision CirrhosisIn development | EchoNet-Liver, NEJM AI 2025 | Retrospective, validated against paired abdominal ultrasound and MRI across two large academic medical centers. |
| Precision ReportingIn development | EchoPrime, Nature 2026 EchoNet-Measurements, JACC 2025 |
Retrospective, validated across five health systems. A prospective blinded randomized trial is in progress. |
| Precision VHDIn development | Circulation 2024 · JAMA Cardiology 2025 | Retrospective, developed at Cedars-Sinai with external validation at Stanford Healthcare. Pipeline detail → |
In progress
Precision Reporting is undergoing a prospective, blinded, randomized clinical trial, and has an FDA Breakthrough Device Designation application submitted.
Precision Cirrhosis holds FDA Breakthrough Device Designation and is enrolled in the FDA's Total Product Lifecycle Advisory Program (TAP).
Precision Cardiac Amyloid is enrolled in FDA TAP and was previously granted Breakthrough Device Designation.
What is not here
There is no prospective randomized trial of Precision Cardiac Amyloid. Both of its validations are retrospective case-control studies. So is the published evidence for the other FDA-cleared echo amyloid device, so this is a parity gap in the category rather than a gap unique to us — but it is the obvious next study, and we would rather you heard it from this page than found it yourself.
Predictive values from a case-control cohort move with the prevalence of the population actually tested. The evidence page shows the full prevalence range rather than a single flattering number.
Talk to our clinical team
For trial design questions, endpoint definitions, or the core lab's role in your own study, our clinical team will take the detail as deep as you want it.