Company Spotlight: GuideAI VAOT - Triage AI for Peripheral Arteries

Sam: When we looked at the schedule today, I was totally ready for an episode about stroke triage. The market for Large Vessel Occlusions in the brain is where everyone focuses. I thought we were looking at a new competitor trying to slash the door-to-needle window.

Maya: It is a fair assumption, but this document throws a fascinating curveball. We are unpacking the 'FDA 510(k) summary K260729' published by the U.S. Food and Drug Administration. Just a quick reminder, our voices are AI-generated, and this is a summary of a public document. And as for your stroke assumption, Sam, this AI is actually looking at the other end of the body.

Sam: Right! We are doing a Company Spotlight on GuideAI and their product, Vascular Assist Occlusion Triage, or VAOT version 1.0. It achieved a sensitivity of 0.915 in testing, and the twist is that it got cleared by comparing itself directly to a stroke AI.

Maya: Exactly. This is a brilliant look at how regulatory pathways actually work for medical software. Let's start with what VAOT actually is. It is a Class II device under Regulation Number 21 CFR 892.2080.

Sam: And the official product code is QAS, which the summary defines as 'Radiological Computer Aided Triage And Notification Software'. But who is it for?

Maya: It is indicated for use in the analysis of CTA studies of adult patients. The goal is to assist hospital networks and trained medical specialists in workflow prioritization.

Sam: So if it is not looking at the brain, what exactly is it looking for?

Maya: The document is very specific. It looks for imaging patterns consistent with possible total arterial occlusion within the aorta, common iliac, external iliac, common femoral, superficial femoral, and popliteal arteries.

Sam: Okay, so we are talking about major vessels in the torso and the legs. Peripheral arterial occlusions. That is a massive shift from intracranial vessels.

Maya: It is. And before we get to the AI performance numbers, we need to talk about how this software actually plumbs into a hospital's IT system. Section 4 outlines the workflow step by step.

Sam: Which is critical for any IT buyer listening. They don't just care if the AI works, they care how it handles their DICOM data. So how does VAOT do it?

Maya: It starts when CTA studies are captured by a CT imaging system and deposited on the PACS. Then, a component called the Gateway instance, which runs acting as a PACS listener, receives a copy of the CTA DICOM study.

Sam: So it is listening on the local network. What does it do with the images?

Maya: The Gateway instance checks the DICOM meta data to confirm it can be processed. Then it de-identifies it, and securely transmits it to the VAOT cloud for processing.

Sam: Meaning the heavy lifting is done off-site in the cloud. The summary says it uses an image-processing pipeline consisting of several deep neural networks.

Maya: Correct. If one or more possible occlusions are flagged by those networks, the cloud selects an axial slice from the original study as a 'preview image' and sends it back to the local Gateway instance.

Sam: And this is where the regulatory guardrails come in. It does not alter the original medical images. That is stated explicitly as a strict rule.

Maya: Exactly. The Gateway instance re-associates that preview image with the patient, embeds it into a DICOM Secondary Capture, and routes it to the DICOM Destination, like the PACS.

Sam: And then the PACS uses that Secondary Capture to notify the radiologist of the suspected occlusion. It is a completely parallel workflow.

Maya: Yes. The summary emphasizes that no cases are removed from the worklist or deprioritized. It operates entirely in parallel with the on-going standard-of-care interpretation workflow.

Sam: Let's dig into the 510(k) pathway itself, because Section 7 on Substantial Equivalence is arguably the most interesting part of this entire document.

Maya: It really is. To get a 510(k) clearance, a company has to prove their device is substantially equivalent to a legally marketed predicate device.

Sam: And the predicate device GuideAI chose is the Aidoc Briefcase LVO Triage, cleared under K220709.

Maya: Which, as you pointed out at the top of the show, is a stroke triage AI. Briefcase LVO Triage analyzes head CTA images to detect complete Large Vessel Occlusion and Medium Vessel Occlusions.

Sam: How does the FDA justify comparing an algorithm trained to look at the brain with an algorithm trained to look at the aorta and the legs?

Maya: The FDA's logic is laid out clearly in the summary. Both devices share the exact same intended use under 21 CFR 892.2080. They are both Class II notification-only tools that analyze CTA studies using deep-learning algorithms.

Sam: The document literally says this represents a difference in indications for use within the same clinical purpose of rapid identification and prioritization of suspected acute vascular occlusion, rather than a new intended use.

Maya: Right. The primary difference is the anatomical location, which requires the algorithms to be trained on a different dataset. But the summary concludes that this difference does not raise new questions of safety and effectiveness.

Sam: That is a massive takeaway for companies building AI. The clinical workflow and the risk profile matter more for equivalence than the specific pixels you are analyzing.

Maya: Exactly. The risks for both devices are defined identically: false positives leading to unnecessary prioritization, or false negatives where the standard of care review simply applies.

Sam: Wait, Maya, I noticed a really interesting discrepancy when reading the comparison table in Section 7 versus the performance testing in Section 8.

Maya: Let's hear it.

Sam: In the comparison table, it lists the notification delivery method. The predicate device uses a 'Standalone Desktop Application'. But VAOT, the subject device, delivers notifications 'Using DICOM Secondary Captures'.

Maya: Yes, we covered that in the workflow. It goes back to the PACS.

Sam: But if you scroll down to Section 8.3, where they define the secondary endpoint for time-to-notification, it says the time includes analyzing and sending a notification 'back to the desktop application'.

Maya: That is a very sharp catch. The document does not explain that discrepancy at all.

Sam: It almost looks like they accidentally pasted language from the predicate device's workflow into their testing description, or maybe they actually used a desktop app just for the retrospective test.

Maya: It is entirely possible, but as we always say, we can only evaluate what is written in the summary. And the summary explicitly states in the substantial equivalence discussion that VAOT relies on DICOM Secondary Captures for its actual cleared workflow.

Sam: Fair enough. Let's talk about that performance testing. How did GuideAI prove that VAOT actually works?

Maya: They conducted a retrospective, blinded, multicenter study. It evaluated the software's performance on 317 cases from US-based clinical sites.

Sam: And were these cases part of their original training data?

Maya: No. The summary specifies that the cases collected for the pivotal dataset were all distinct from the cases used to train the algorithm.

Sam: Good. You cannot grade your own homework. Who decided what the 'right answer' was for these 317 cases?

Maya: The ground truth was determined by 3 expert US board certified Radiologists, using majority voting.

Sam: Out of the 317 cases, the breakdown was 165 positive for occlusion, and 152 negative.

Maya: Which means the study sample was enriched. In a normal clinical queue, you wouldn't expect more than half the scans to be positive. The summary explicitly notes the set was enriched with 21 positive vessel aortic occlusions cases.

Sam: What about the patient demographics?

Maya: It covered a broad range of patient ages, and the gender distribution was 180 Male and 137 Female.

Sam: And the hardware? AI can be notoriously fragile when you switch scanner brands.

Maya: Section 8.4 provides a frequency distribution of the manufacturers. GE MEDICAL SYSTEMS was the largest bucket, with 192 cases, which is 60.6%.

Sam: And the rest of the market?

Maya: SIEMENS had 73 cases, or 23.0%. Canon/Toshiba had 37 cases at 11.7%. Philips had 14 cases at 4.4%. And there was 1 case listed as Other, making up 0.3%.

Sam: So it is heavily weighted toward GE, but it has seen data from all the major vendors. So, what were the actual results? Did they hit their marks?

Maya: They did. VAOT met the primary endpoints of sensitivity and specificity, which had an 80% performance goal.

Sam: The sensitivity was 0.915, with a 95% confidence interval of 0.862 to 0.953.

Maya: And the specificity was 0.901, with a 95% confidence interval of 0.842 to 0.944.

Sam: That is strong performance, especially on a retrospective dataset. What about speed? You mentioned a secondary endpoint earlier.

Maya: The secondary endpoint was a time-to-notification of less than 5 minutes. The summary states that VAOT met that endpoint.

Sam: Which is exactly what you need for a triage tool. It has to flag the study before the radiologist naturally gets to it in the queue. But what trips this algorithm up? What are the confounders?

Maya: The document is very transparent about this. In the negative cases, they list specific pathologies present that acted as confounders. They call out 'Non-occlusive arterial disease (mimics)'.

Sam: So, disease that looks bad but isn't fully blocking the vessel.

Maya: Exactly. They also list 'Post-operative and device confounders', 'Out-of-scope vessels/locations', 'Other vascular pathologies', and 'Acquisition/technical quality issues'.

Sam: That is the reality of clinical imaging. Patients have stents, they have previous surgeries, or they just move in the scanner. The data is messy.

Maya: Which brings us directly to the limitations and warnings in the document. The FDA clearance is strictly for triage and notification. VAOT is not intended for primary diagnosis, screening, disease detection, or patient management decisions.

Sam: The wording in Section 5 is incredibly direct. It says 'VAOT does not detect, localize, characterize, measure, or diagnose vascular disease'.

Maya: It sounds contradictory to a layperson because the tool is literally looking for occlusion patterns, but legally and medically, it does not diagnose.

Sam: The notifications might include compressed preview images, but the summary stresses these are provided for informational purposes only and are not intended for diagnostic use.

Maya: The interpreting clinician remains responsible for viewing the full imaging study according to standard clinical practice.

Sam: Are there any specific patients this shouldn't be used on?

Maya: Yes, Section 6 lists a contraindication for use: The VAOT device was not tested on pediatric patients.

Sam: And again, this is a 510(k) clearance based on substantial equivalence. The FDA letter from Jessica Lamb reminds the company that this determination does not mean the FDA has made a determination that the device complies with other requirements of the Act.

Maya: Exactly. Things like the Quality Management System Regulation, or QMSR. Section 8.5 notes they completed software verification and validation, and Section 8.6 notes they conducted cybersecurity risk assessment and testing.

Sam: Those are the baseline hurdles for any hospital IT buyer. You have to prove you won't expose patient data when you route those images to the cloud.

Maya: Ultimately, GuideAI concludes in Section 9 that the introduction of VAOT into the U.S. market should not raise any safety or efficacy concerns, as it shares the same intended use and principles of operation as the predicate device.

Sam: It is a masterclass in regulatory strategy. Using a brain stroke AI to clear a lower extremity AI because the clinical triage workflow is functionally identical. That is a memorable takeaway.

Maya: It certainly is. To recap, we unpacked the 510(k) summary for GuideAI's VAOT version 1.0, a cloud-based triage tool for lower extremity and torso arterial occlusions.

Sam: It hit a sensitivity of 0.915 on a retrospective dataset of 317 cases and operates in a completely parallel workflow using DICOM Secondary Captures.

Maya: You can read the full FDA document for yourself. We have linked it in the show notes.

Sam: And a final reminder: this podcast is an analysis of a public regulatory document, not medical advice. Thanks for listening.