Seven questions to ask any amyloid AI vendor.
Two cleared devices can report very different numbers and both be working exactly as designed. The figure that matters is rarely the one on the front of the brochure. These are the questions that separate them — including the ones we would rather you asked us.
What is the specificity — and what does that cost per 1,000 studies?
Ask for specificity, not just sensitivity or AUC, then convert it. A device at 99.0% specificity produces about 10 false flags per 1,000 studies. A device at 89.7% specificity produces about 100. Every false flag costs a workup — bone-tracer scintigraphy, immunofixation, free light chains, sometimes biopsy. Ask the vendor to state the false-flag rate per 1,000 studies at your prevalence.
What is the exact intended-use population?
The indication for use is the sentence that governs which of your patients the device may be used on, and the differences between devices are large. Some clearances require a heart failure context, or a prior clinical suspicion, before the device may run. A narrower population means a smaller share of your echo volume is eligible, and it means someone has to decide which studies qualify before the algorithm ever sees them. Ask for the indication verbatim from the clearance — the age range, and any required clinical context — and work out what fraction of your annual studies actually falls inside it.
How many views does it read?
Some devices read a single apical four-chamber view. Others read apical four-chamber and parasternal long-axis. The amyloid phenotype is not confined to one window, so ask how many views the model uses and what happens when a required view is missing or unusable.
Which amyloidosis types were validated?
AL and ATTR are different diseases with different urgency. Untreated AL amyloidosis has a survival horizon measured in months and needs hematology, not cardiology. Much published evidence in this category is ATTR-weighted. Ask which types were represented in the validation cohort and in what proportion.
Was performance published by site?
A pooled figure across sites hides the range. Ask for performance at every validation site, including the worst one. A vendor willing to publish its floor is telling you something a pooled number cannot. Also ask how cases and controls were selected — an AUC is a property of a model and a cohort, so two AUCs from differently constructed cohorts are not comparable.
Has an independent group audited it for bias?
Ask whether any group outside the vendor has evaluated the model for bias, and ask to see the metric that failed as well as the one that passed. Equal opportunity and demographic parity ratios are the standard measures. A vendor that can only show you favourable audit results has not shown you an audit.
Can the result arrive before sign-off?
A result that lands after the cardiologist has signed the report is a result that changes nothing on this study. Ask when the output arrives relative to sign-off, measured from study completion rather than from the algorithm receiving its last frame — those are very different numbers in a cloud deployment. Ask for the median, the 75th percentile, and the maximum, and ask what happens to studies where a required view is missing.
Our answers, for the record
The same seven questions, answered for InVision Precision Cardiac Amyloid. Every figure is sourced on the evidence page.
| Question | InVision Precision Cardiac Amyloid |
|---|---|
| Specificity | 99.0% in FDA validation (K243866); 0.98 in an independent head-to-head. About 10 false flags per 1,000 studies. |
| Intended population | Adults aged 65 and older — with no requirement for heart failure, prior suspicion, or any other clinical context. Every routine study in that age range is eligible. |
| Views | Two: apical four-chamber and parasternal long-axis. |
| Types validated | Both. In the international validation the case mix was ATTR 74.2%, AL 22.7%. |
| Site-level results | Published, including the lowest: 0.833 at Kumamoto University, up to 0.944 at Keio University. |
| Independent bias audit | Yes, in a Northwestern-led study. Equal opportunity ratio 1.05 for patients who identified as Black. We also publish the metric that was flagged: demographic parity 0.64 for Hispanic patients. |
| Turnaround | Median 0 minutes from study completion — the result is typically available before the study finishes. 75th percentile 2 minutes, maximum 10. It reaches the reader before sign-off, not after. Turnaround detail → |
| Where we score lower | Sensitivity 0.57 and false negative rate 0.43 in the same independent study. We publish the whole table. |
Run these questions past us
Bring the list to a call with our clinical team, or send it to us in writing and we will answer in writing.