Company Spotlight: Anumana WatchMate - Turning 2D Echo into 3D Surgical Plans

Maya: What if an AI could take a standard 2D ultrasound of your heart, reconstruct it into a 3D model, and beat a panel of expert physicians at measuring a surgical target by 2.1 mm?

Sam: Today we are looking at exactly that. This is the FDA 510(k) summary for K260300, clearing the WatchMate Software by Anumana, published on June 15, 2026. Just a quick reminder right at the top: our voices are AI-generated, and this is a summary of a publicly available regulatory document, not promotional material from the company.

Maya: This is a really fascinating Company Spotlight episode because it touches on a very specific, high-stakes procedure in cardiology, and shows exactly how machine learning is being evaluated by the FDA to assist in the cath lab.

Sam: Let's start with the basics. What is the WatchMate Software, and what is it actually cleared to do according to this document?

Maya: WatchMate Software is a standalone software application. It is meant to be used by healthcare professionals during left atrial appendage closure procedures, which are often abbreviated as LAAC.

Sam: Right, and for those who might not be in the cardiovascular space, the left atrial appendage is essentially a small pouch in the top left chamber of the heart. In some patients, blood can pool there and form clots, leading to strokes. So they go in and plug it up.

Maya: Exactly. But to plug it, you have to know exactly what size and shape the pouch is. According to the indications for use, WatchMate processes transesophageal echocardiography images, or TEE. Those are ultrasound images taken from inside the esophagus, right behind the heart. The software assists in pre-procedural planning and device sizing.

Sam: And the key feature here is that it takes those 2D TEE images and provides a 3D reconstruction of the appendage. It also provides automated quantitative measurements of the dimensions and the landing zone parameters. But the document is very careful to state that it functions as an adjunctive tool. It provides anatomical insights without replacing clinical judgment.

Maya: That is standard language for this type of software, but crucial for hospital liability. Speaking of which, the FDA letter specifically notes that the Center for Devices and Radiological Health does not evaluate information related to contract liability warranties. The clearance is strictly about safety and effectiveness based on substantial equivalence.

Sam: Let's dig into how this AI actually works. The summary details a multi-stage deep learning pipeline, and I found the breakdown fascinating. It is not just one big black box algorithm.

Maya: No, it is highly structured. There are distinct models working in sequence. First, there are video validation models that classify the 2D input and confirm the left atrial appendage is actually in the frame.

Sam: Then comes stage two, which I thought was a really clever engineering workaround. It uses character recognition models to extract the transducer angle and imaging depth metadata directly from each frame.

Maya: Right, they are literally having the AI read the text off the ultrasound screen to understand the physical orientation of the probe. Stage three is the anatomy segmentation model. It identifies the left atrial appendage, the left atrium, pulmonary vein, and left ventricle.

Sam: Stage four takes all those 2D contours and builds a densified 3D point cloud, converting it into a smooth surface mesh. And finally, stage five is the landing zone estimation algorithm, which gives the clinician a tentative plane and quantitative measurements for where the closure device should sit.

Maya: As an analyst, what I immediately look for is the training data. Where did these models come from? The document states they were developed using a de-identified clinical dataset sourced from a large United States health system, spanning over 35 years of imaging data.

Sam: And they give us the exact numbers on that data. For the anatomy segmentation, they trained on 36,541 annotated TEE frames, and verified on 7,385 frames.

Maya: But for the 3D generation, they had to map the ultrasound to actual 3D ground truth. So they trained the 3D models on 61,284 cardiac CT scans, and verified on 3,270 CT scans. They note that the data was split at the patient level to prevent data leakage.

Sam: That is a massive training set, but I have to point out what is not there. It says a large United States health system, singular. It does not name the system, and it implies all this retrospective training data came from one institutional source. In the AI world, we always wonder how well a model generalizes to other hospitals with different equipment and patient demographics.

Maya: That is a very valid point, and it leads perfectly into how they proved it actually works. Because the FDA clears these devices based on substantial equivalence, WatchMate had to be compared to a predicate device.

Sam: The primary predicate device they chose is the TruPlan CT Imaging Software, cleared under K222593. And there is a really interesting technical leap here. TruPlan uses CT scans as its input data. WatchMate uses TEE ultrasound as its input data.

Maya: Exactly. They are using different modalities to reach the same clinical goal: adjunctive, clinician directed Left Atrial Appendage Closure planning. The FDA accepted this because, as the document states, the differences in modality do not raise new or different questions of safety and effectiveness.

Sam: To prove that their ultrasound-based AI was as good as the CT-based standard, Anumana ran a prospective, multi-center clinical data collection study. The trial number is NCT07126600.

Maya: This is where the numbers get really interesting. They enrolled 100 subjects across 6 U.S. centers. The goal was to see if WatchMate, looking at standard intra-procedural 2D ultrasound, could measure the landing zone as accurately as humans reading the ultrasound, when compared to a gold standard of a CT scan.

Sam: And how did they define that gold standard?

Maya: They used three independent expert cardiac CT readers to establish a structured unanimous-consensus baseline. Basically, three experts looking at high-resolution 3D CT scans had to agree on the true measurements. Then, they compared WatchMate to the average of three blinded physicians who were interpreting the 2D ultrasound.

Sam: Okay, so it's the AI versus the human average, both trying to match the CT scan reality. What were the results?

Maya: For the primary endpoint, the mean absolute deviation versus the CT scan for WatchMate was 3.1 ± 3.4 mm.

Sam: And for the physicians reading the 2D ultrasound?

Maya: It was 5.2 ± 3.4 mm. So WatchMate had a difference of -2.1 mm compared to the human readers.

Sam: Wow. So the AI was closer to the CT ground truth by 2.1 mm.

Maya: Yes, the document states it met the pre-specified non-inferiority margin of 1 mm, with a P value of less than 0.001. And because the upper bound of the 97.5% confidence interval was less than zero, it additionally met superiority.

Sam: That is a huge win for the software. But I want to go back to the patient numbers for a second. You said they enrolled 100 subjects. But the document says only 72 were included in the per-protocol population.

Maya: Good catch. The document states that 100 subjects were enrolled, with 72 included in the per-protocol population. Exclusions occurred due to unanalyzable TEE data or WatchMate 2D-3D inconsistency flags.

Sam: That feels significant. If I am a hospital administrator looking to buy this, or a clinician relying on it, that drop in the study population means the software might just throw its hands up and refuse to analyze a large chunk of patients.

Maya: That is exactly what it means in practice. The AI has built-in quality control, which is good for safety. If the ultrasound image is poor, or if the 2D to 3D mapping doesn't make sense, it flags it and stops. But from a workflow perspective on the ward, it means you can't guarantee this tool will work for every single patient you wheel into the electrophysiology lab.

Sam: Which probably explains why the FDA stresses that it does not replace clinical judgment. It's an adjunctive tool. If it flags out, the physician still has to do their job the traditional way.

Maya: Precisely. But when it does work, the Bland-Altman analyses in the study showed minimal systematic bias for WatchMate, whereas the 2D TEE readers demonstrated consistent underestimation. So humans tend to measure the hole as smaller than it actually is.

Sam: Let's pivot a bit to how this actually gets installed in a hospital, because the IT requirements are always a nightmare for these AI tools. The summary outlines the system components pretty clearly.

Maya: It does. WatchMate runs on a dedicated Linux workstation with an NVIDIA GPU of 24 gigabytes or more. And the user interface is on an iPad, which they call the Mobile Platform. It also supports an optional HDMI display so the whole cath lab can see it.

Sam: What stood out to me is the interoperability and cybersecurity testing. It operates on a local private network. It pulls the DICOM images straight from the ultrasound console via Ethernet or Wi-Fi, and it's validated only for compatibility with GE and Philips TEE ultrasound systems.

Maya: That local-only network piece is huge for the boardroom. Many AI tools are cloud-based, which requires passing sensitive patient data through hospital firewalls out to a vendor's server. WatchMate does the heavy lifting locally on that Linux box.

Sam: Right, the document specifically calls out features like LUKS full disk encryption, mutual TLS, and a local-only network. They even did a Mock Environment Cybersecurity Test to verify resilience against threats.

Maya: It shows a maturity in how they are approaching hospital IT procurement. If you want a hospital CIO to approve your device, a local, encrypted, closed-loop system is generally an easier sell than a cloud dependency.

Sam: Let's talk about the reference devices mentioned. Beside TruPlan, they listed Philips 3D Auto LAA and the Nesa Medtech Fibroid Mapping Reviewer Application. Why include a fibroid mapping tool for a heart device?

Maya: It is a strategy for 510(k) submissions. The Nesa Medtech device takes 2D ultrasound images and generates a 3D model of the uterus. By citing it as a reference device, Anumana is demonstrating to the FDA that the underlying principle of taking 2D ultrasound and turning it into a 3D measurable model is already a cleared concept in other areas of medicine.

Sam: That makes a lot of sense. You piece together the regulatory puzzle. TruPlan proves we can do LAAC planning with software. Philips proves we can do it on an ultrasound cart. And Nesa proves we can post-process 2D into 3D. Put it all together, and you get WatchMate.

Maya: Exactly. And the summary concludes that the differences in input modality, workflow, and platform do not raise new questions of safety or effectiveness. That is the magic phrase for a 510(k) clearance.

Sam: So, Maya, bringing it all together: what is the one memorable takeaway from this document for a clinician on the ward, and for a health-system leader in the boardroom?

Maya: For the clinician, the takeaway is that AI can now take the standard 2D echo you are already doing in the procedure room and build a 3D plan that is technically closer to a CT scan than human estimation is. But, you have to be prepared for the software to reject scans if the image quality isn't perfect or there are inconsistency flags.

Sam: And for the companies building or buying AI?

Maya: The takeaway is that the FDA is entirely comfortable clearing complex, multi-stage deep learning pipelines, provided you can benchmark them against an established clinical ground truth, like a multi-reader CT consensus. Also, edge computing—running heavy AI locally on a dedicated GPU workstation—is a viable path for regulatory clearance and hospital deployment.

Sam: It is a great example of where the field is heading. Closing the gap between a standard, lower-cost imaging modality like ultrasound, and a high-resolution, high-cost modality like CT.

Maya: Absolutely. Just remember that it is cleared for adult patients only, and requires trained healthcare professionals to review and approve the landing zone estimation before any device is actually deployed.

Sam: That brings us to the end of this deep dive. To recap, we've been looking at the FDA 510(k) summary for Anumana's WatchMate Software, a tool that uses AI to turn 2D TEE images into 3D plans for closing the left atrial appendage. It beat human averages in clinical testing but requires a clean signal to work.

Maya: As always, you can find a link to the full public FDA document in our show notes. We encourage you to read the performance data and trial design for yourself.

Sam: And a final reminder: this podcast is for informational purposes only. It is not medical advice, and we do not endorse any specific products or companies. Thanks for listening.