Explainer

How AI detects cardiac amyloidosis on an echocardiogram.

Cardiac amyloidosis is treatable, but it is usually found late because its signs hide in plain sight on a routine echo. AI can read that same echo and flag the studies that deserve a closer look — at the moment the study is read, with no new test.

What cardiac amyloidosis is, and why it hides

Cardiac amyloidosis is an infiltrative disease in which misfolded protein deposits stiffen the heart muscle, producing restrictive cardiomyopathy and heart failure. It comes in two main forms — transthyretin (ATTR) and light-chain (AL) — and both are now treatable, so finding it sooner changes outcomes.

It is missed because its features overlap with ordinary heart failure, and the clues on an echo are easy to read past. Systematic screening keeps turning up far more disease than is recognized in routine care: among hospitalized patients aged 60 and older with heart failure and a wall thickness of 12 mm or more, 13.3% had ATTR cardiac amyloidosis (González-López, Eur Heart J 2015), and about 1 in 6 patients undergoing TAVR for aortic stenosis had it on dedicated imaging (Castaño, Eur Heart J 2017). Patients are often diagnosed years after their first cardiac symptoms.

What the echo shows, and why it is easy to miss

The clues to amyloid are already on a standard echo. They are subtle and only form a pattern when several are read together, so a single one is easy to pass over. These are the same structures the AI reads.

Thick walls, small cavity

A left-ventricular wall of 12 mm or more with no other cause, such as hypertension or aortic stenosis; a small cavity with a high relative wall thickness (a thick wall for the chamber size); and a granular, sparkling look to the muscle — a soft sign, less reliable with modern harmonic imaging.

The wall thickens with protein

Large atria, thick atrial septum

Both atria enlarged under high filling pressure; a thickened interatrial septum and valve leaflets; a right-ventricular free wall above 5 mm; and, in about 30% of patients, a small pericardial effusion.

Infiltration spreads

The strain pattern

On longitudinal strain the base gives out first while the apex keeps contracting — relative apical sparing, the "cherry on top." The ejection fraction can read near-normal while long-axis function is already poor.

Base fails, apex spared

Outside the heart, the pattern often shows years earlier: carpal tunnel syndrome in both wrists, 5 to 9 years before; an ECG voltage that is low for how thick the wall is; blood pressure that falls, so long-standing hypertension eases or heart-failure drugs are no longer tolerated. Suspect cardiac amyloidosis in heart failure with a thick wall, a non-dilated ventricle, and no clear cause.

For the full illustrated version — the misfolded protein, the tissue, the remodeling and every echo view — see The Amyloid Heart, an interactive tutorial from InVision.

How the AI reads the echo

The analysis runs on a study the lab has already performed — no new acquisition, no hardware, no change to the scan.

Standard views in

It reads two views from a routine transthoracic echo: apical-4-chamber (A4C) and parasternal-long-axis (PLAX). These are already captured in a normal study.

No new test

Two signals

A model measures left-ventricular wall thickness from the PLAX view; a second model assesses tissue texture and motion from the A4C view — a pattern present in the images but not reliably perceptible to the eye.

Structure + texture

Flag out

The result returns into the originating study as a DICOM secondary capture, available to the reader as the last frame — so it surfaces inside the normal reading workflow.

In the read

The algorithm is static and locked: every inference is tagged to the cleared model version, and the model cannot change without a new FDA clearance. The underlying method was peer-reviewed in JAMA Cardiology (2022) and validated in JACC: Advances (Duffy, 2025).

What the result means

InVision Precision Cardiac Amyloid is decision-support AI. It flags echoes suggestive of cardiac amyloidosis and alerts the interpreting physician for referral to confirmatory testing. It does not diagnose, and patient-management decisions should not rest on its result alone.

A flag is the start of a work-up, not a conclusion. The value of a high-specificity flag is that it is far more likely to be real, so the confirmatory pathway is spent on the patients most likely to have the disease.

After a flag

  1. Read the clues. Check the history and the ECG for the signs above — carpal tunnel syndrome, a QRS voltage low for the wall thickness, blood pressure that has fallen.
  2. Confirm. Two tests settle most cases: a monoclonal-protein screen (serum free light chains with the κ/λ ratio, serum and urine immunofixation) and a bone scan with SPECT.
  3. Refer the urgent form. If a monoclonal protein is present, refer to hematology without delay — the light-chain (AL) form moves fast, with a median survival of 5.8 months once it reaches the most advanced stage.
The software flags. The physician authors. No InVision output is a diagnosis. The interpreting physician reviews the flag with the study and the patient's history, and decides on referral.

How accurate it is

In the FDA 510(k) validation, specificity was 99.0% and sensitivity was 60.7% across 1,221 studies at three US sites. That is the highest specificity and the highest positive predictive value of any FDA-cleared echo AI for cardiac amyloidosis, so at the low prevalence of undiagnosed disease in a real echo lab a flagged study is far more likely to be amyloid. Positive predictive value depends on how common the disease is in the population assessed.

Why specificity carries the weight: most echoes have no amyloid. At a 1% prevalence, 990 of every 1,000 studies are negative, so each percentage point of specificity is about 10 fewer false flags per 1,000 echoes — and every false flag sends a patient to a bone scan and blood and urine tests they did not need. At that prevalence about 2 in 5 flagged studies are amyloid (PPV 38.4% at 1% prevalence). The trade is sensitivity: a flag appears on about 6 in 10 amyloid studies, so a study without a flag does not rule amyloid out.

It is also the only FDA-cleared echo-AI amyloid indication with no heart-failure and no wall-thickness prerequisite: it applies across adults aged 65 and older undergoing echocardiography. See the full comparison, or the complete evidence, including where the model scores lower.

FAQ

Frequently asked questions

Does AI diagnose cardiac amyloidosis from an echocardiogram?

No. InVision Precision Cardiac Amyloid is decision-support AI; no InVision output is a diagnosis. It flags echoes suggestive of cardiac amyloidosis so the interpreting physician can refer the patient for confirmatory testing. The physician makes the diagnosis.

What echocardiographic views does it analyze?

Two standard views from a routine transthoracic echo: apical-4-chamber (A4C) and parasternal-long-axis (PLAX). No additional acquisition, hardware, or special protocol is required.

How is it different from reading the echo the usual way?

It runs on studies the lab has already acquired and surfaces a signal that is present in the images but not reliably perceptible to the eye, so suspicion is raised at the moment the echo is read rather than years later. It does not change acquisition and adds no scan time.

Is it FDA cleared?

Yes. FDA 510(k) K243866, cleared 21 May 2025 (product code SDJ, Class II, prescription use), for adults aged 65 and older undergoing echocardiography.

How accurate is it?

In the 510(k) validation, specificity was 99.0% and sensitivity was 60.7% (1,221 studies, three US sites) — the highest specificity and positive predictive value of any FDA-cleared echo AI for cardiac amyloidosis. Positive predictive value depends on prevalence.

What does cardiac amyloidosis look like on an echocardiogram?

The red flags are a left-ventricular wall of 12 mm or more with no other cause, a small cavity with a high relative wall thickness, both atria enlarged, a thickened interatrial septum and valve leaflets, and often relative apical sparing on longitudinal strain — the base gives out while the apex keeps contracting. The signs are subtle and overlap with ordinary heart failure, which is why they are easily read past.

What should a clinician do after an AI flag for cardiac amyloidosis?

Treat a flag as the start of a work-up, not a conclusion. Review the history and ECG for supporting clues (carpal tunnel syndrome, a QRS voltage low for the wall thickness, blood pressure that has fallen), then confirm with a monoclonal-protein screen and a bone scan with SPECT. If a monoclonal protein is present, refer to hematology without delay for the light-chain (AL) form.

InVision Precision Cardiac Amyloid (FDA 510(k) K243866) is decision-support software for adults aged 65 and older undergoing echocardiography. It flags studies suggestive of cardiac amyloidosis for referral to confirmatory testing; it does not diagnose, and patient-management decisions should not rest on its result alone. 510(k) validation: sensitivity 60.7%, specificity 99.0% (1,221 studies, 3 US sites). PPV depends on prevalence (510(k) table: 38.4% at 1%). Highest specificity and PPV: comparison of the FDA 510(k) summaries of FDA-cleared echo-AI devices for cardiac amyloidosis; cross-study, different study populations; PPV compared at equal prevalence. Echo red flags follow the 2021 ESC position statement on cardiac amyloidosis and the ASE/EACVI chamber-quantification standard (Lang 2015). As of October 2026. Prescription use only.