Company Spotlight: Caristo CaRi-Plaque - 0.999 Correlation with Remote Operators

Sam: Imagine launching a new diagnostic software that achieves a 0.999 correlation with expert cardiologists for measuring calcified plaque, but doing it with threshold-based tools and remote human operators instead of on-site deep learning.

Maya: That is exactly what we are looking at today. We are unpacking the FDA 510(k) summary for a device called CaRi-Plaque, manufactured by Caristo Diagnostics, which the U.S. Food and Drug Administration published on February 20, 2025.

Sam: Before we dive in, a quick reminder that our voices are AI-generated. This is a Company Spotlight episode, but we want to be completely clear: this is a summary of the FDA's public 510(k) summary document, not a summary of company marketing material.

Maya: Exactly. We only know what the FDA document tells us. We do not have any external information about the company's funding, their customer base, or their other commercial products. We are focusing strictly on what the agency cleared and the real numbers behind that clearance.

Sam: Right. So let us start with the basics. What exactly is CaRi-Plaque, and what is its intended use?

Maya: The document defines CaRi-Plaque version 1.0 as a web-based, software-only application. It is a Class II device under the product code LLZ, which stands for Medical Image Management And Processing System.

Sam: And what does it process?

Maya: It takes previously acquired Computed Tomography, or CT, Angiographic DICOM images. Its purpose is to provide quantitative and qualitative clinical analysis to characterize and quantify plaque formation and stenosis in coronary arteries.

Sam: Okay, so this is for cardiologists and radiologists looking at heart scans to see how blocked or diseased the arteries are. But the workflow described here is quite different from what we usually see with local workstation software.

Maya: It really is. The summary explicitly states that CaRi-Plaque is intended for use by internal operators who have been appropriately trained in the software's functions, capabilities and limitations.

Sam: Internal operators? So the hospital is not running this themselves?

Maya: Correct. The hospital transmits the DICOM imaging data via a secure picture archiving and communication system, or PACS. Then, trained operators located in regional data centers use the web-based software to perform the analysis. They generate a CaRi-Plaque Report, which is subsequently provided to the healthcare professional end user.

Sam: That is a massive workflow distinction for any health system leader listening. You are essentially outsourcing the post-processing of the scan. You do not need to install heavy software on your own desktop computers or train your own staff to use it.

Maya: Yes, but it also means you are relying on the quality and turnaround time of those remote operators. And the technology they are using is interesting, too. The software uses a combination of 3D image thresholding computerized algorithms and manual editing tools to quantify the atherosclerotic plaque.

Sam: Manual editing tools. So it is very much a human-in-the-loop system. The operator is actively defining and correcting the vessels of interest. Let us compare this to the predicate device. How does a company prove to the FDA that this is safe and effective?

Maya: Through the 510(k) pathway, a company must prove its device is substantially equivalent to a legally marketed predicate device. In this case, Caristo used Autoplaque 3.0 from Cedars-Sinai Medical Center.

Sam: The document has a whole section comparing the two. What are the major differences?

Maya: The FDA outlines three key technological differences. First, as we discussed, CaRi-Plaque is web-based, while the predicate resides on a local commercial computer platform. Second, the operator for CaRi-Plaque is at a remote site, while the predicate operator is at the clinical site.

Sam: And the third difference?

Maya: The third is the algorithmic approach. CaRi-Plaque uses threshold-based segmentation, whereas the predicate device uses machine learning to train the algorithms.

Sam: Wait, really? They moved away from machine learning to a threshold-based method? That feels counterintuitive given all the hype around AI right now.

Maya: It might seem that way, but the FDA summary notes that this technological methodology difference does not raise questions of differences in safety or effectiveness. Both devices are intended to accomplish the same indicated uses and were verified using the same clinical testing approach.

Sam: Which means comparing the software's outputs to what expert human readers determine is the ground truth. What exactly does the final report measure?

Maya: The report includes several quantitative outputs. Specifically, it provides Plaque Burden, Calcified Plaque Volume, Total Plaque Volume, Noncalcified Plaque Volume, Low Density Noncalcified Plaque Volume, Remodeling Index, and Maximum Stenosis.

Sam: Those are the core metrics for evaluating coronary disease. So how did they prove this threshold-based, remote-operator system works in practice? Let us get into the clinical validation study.

Maya: They conducted a clinical performance test on a multi-center, international patient population consisting of 117 subjects.

Sam: Just 117 subjects?

Maya: Yes, 117. And the demographics are quite specific. There were 85 men and 32 women, aged 27 to 85 years of age. They collected data from four sites in total—2 US and 2 outside the US.

Sam: That is a very small dataset for a validation study, and a huge gender imbalance. 85 men to just 32 women. Does the document explain why the cohort was so heavily male?

Maya: It does not. It simply reports the numbers. And the way they handled ethnicity data is also highly unusual and worth noting for anyone evaluating the robustness of this data.

Sam: What do you mean?

Maya: They only collected individual subject-level ethnicity at two of the four study sites. That covered 57 subjects. For those 57, the breakdown was 76% White, 4% Asian, 1% Middle Eastern, 14% other, and 5% unknown.

Sam: So for the remaining subjects where they did not collect ethnicity directly, how did they report it?

Maya: The document states that for the two remaining sites, ethnicity was estimated from local population census data.

Sam: Wait. They just guessed the patients' ethnicities based on the census data of the area where the hospital is located?

Maya: That is exactly what the summary says. For the first of those sites, which had 48 subjects, they applied the census breakdown of 44% White, 22% Black, 27% Hispanic or Latino, and 5% Asian. For the second site with 12 subjects, they estimated 46% White, 19% Black, 26% Hispanic or Latino, and 13% Asian.

Sam: That is a major limitation. If you are a hospital buyer, you have to acknowledge that the demographic validation here is partly estimated, not strictly recorded. More than half the dataset's ethnicity is a statistical guess.

Maya: It is a limitation. But to be balanced, the document does state that subgroup analysis by vessel, sex and age did not reveal any bias in performance.

Sam: Alright, let us look at that performance. How did the software's measurements compare to the ground truth set by the independent medical experts?

Maya: They measured agreement using Pearson Correlation Coefficients. The results were exceptionally high. For Calcified Plaque Volume, the correlation was 0.999.

Sam: 0.999 is essentially perfect correlation.

Maya: It is. Total Plaque Volume was 0.985. Noncalcified Plaque Volume was 0.977. Maximum Stenosis was 0.930. Plaque Burden was 0.885, and Low Density Noncalcified Plaque Volume was 0.817.

Sam: Those are incredibly strong correlations. What about the Remodeling Index?

Maya: That was assessed differently, using Cohen's Kappa. The agreement was 63.3%, which gave a Kappa value of 0.42.

Sam: A Kappa of 0.42 is generally considered just moderate agreement. Why was that so much lower than the volumetric measurements?

Maya: The summary does not explain why. It only notes that these results met the predetermined acceptance criteria.

Sam: Okay. They also broke this down by the specific coronary arteries, right? The right coronary artery, the left anterior descending, and so on. Let us walk through those.

Maya: Yes, they tested agreement at the branch level. Let us start with the right coronary artery and its branches, the RCA, R-PDA, and R-PLB. Total Plaque Volume correlation was 0.958. Calcified Plaque Volume was 0.964. Maximum Stenosis was 0.863.

Sam: Still very solid. What about the left anterior descending, the LAD?

Maya: For the LAD and its branches, the D1, D2, and Ramus, it was even better. Total Plaque Volume was 0.959, and Calcified Plaque Volume was 0.997. Maximum Stenosis was 0.929.

Sam: And the left circumflex artery, the LCX?

Maya: Here is where it gets a bit complex based on how the FDA document is written. For the LCX, OM1, and OM2 branches, they list Maximum Stenosis, Total Plaque Volume, Calcified Plaque Volume, Noncalcified Plaque Volume, Plaque Burden and Remodeling Index. That is six metrics.

Sam: Okay, six metrics. What were the numbers?

Maya: The document lists seven numbers for those six metrics: 0.911, 0.948, 0.991, 0.934, 0.888, 0.633, and 90.3%.

Sam: Wait, seven numbers for six metrics? It looks like there is a typo in the FDA text itself. If you look at the previous artery groups, the 0.633 value was the correlation for Low Density Noncalcified Plaque Volume. But they left that metric out of the text for the LCX group.

Maya: You are absolutely right. The very next sentence states: There was insufficient low density non-calcified plaque in the LCX, OM1 and OM2 to generate a meaningful correlation.

Sam: So they didn't have enough data in that specific branch for that specific type of plaque, which makes sense with only 117 subjects. But whoever drafted the summary accidentally left the numerical placeholder in the list. That is a great catch.

Maya: It highlights why we read these documents so closely. Despite the typo, the overall correlations for the LCX were still very high, like 0.991 for Calcified Plaque Volume.

Sam: Before we wrap up, we need to talk about the risks and the specific warnings on the label. This is crucial for the clinicians actually looking at these reports.

Maya: Yes. The indications for use section contains a very specific warning about how the images are rendered. It states: Users should be aware that certain views make use of interpolated data. This is data that is created by the software based on the original data set.

Sam: So the software is filling in the gaps to create a smoother or more complete 3D image. What is the danger there?

Maya: The FDA summary is explicit: Interpolated data may give the appearance of healthy tissue in situations where pathology may be present that is near or smaller than the scanning resolution.

Sam: That is a huge caveat. If a piece of plaque is incredibly small—smaller than the resolution of the CT scanner itself—the software's interpolation might smooth it out, making the vessel look perfectly healthy.

Maya: Exactly. This is why the document stresses that the analysis results are not intended to replace the skill and judgment of a qualified medical practitioner. The physician must review the results alongside the patient's original CT images, clinical history, and other tests.

Sam: And what happens if the physician looks at the original images and thinks the software's interpolation missed something, or they just disagree with the remote operator's analysis?

Maya: The workflow accounts for that. The document notes that the healthcare professional may request a re-analysis of the CT scan data if they do not agree with the report analysis.

Sam: That makes sense, especially since the initial review is done by an internal operator at Caristo's regional data center. Speaking of those operators, did the study test whether different operators get the same results on the same scan?

Maya: It did. The summary mentions that there was good inter-operator agreement and repeatability at the branch level within the CaRi-Plaque device arm. However, it does not provide the specific numerical data for that inter-operator agreement.

Sam: This whole summary paints a really interesting picture of modern medical device clearance. You have a service-based workflow, remote operators, threshold-based algorithms instead of deep learning, a small validation set of 117 subjects with some estimated demographics, and yet, they proved substantial equivalence by hitting a 0.999 correlation against expert ground truth.

Maya: It shows that in the FDA's framework for a 510(k), the underlying technological methodology matters less than whether you can prove the final output is as safe and effective as the predicate device.

Sam: Alright, let us recap. Today we looked at the FDA 510(k) summary for CaRi-Plaque by Caristo Diagnostics. It is a web-based, Class II radiological image processing system cleared to characterize and quantify coronary plaque and stenosis.

Maya: It relies on internal trained operators at regional data centers using threshold-based software, distinguishing it from local machine-learning predicates. While the clinical validation dataset was small and had demographic limitations, it demonstrated exceptional correlation with expert readers, including a 0.999 Pearson Correlation for Calcified Plaque Volume.

Sam: But clinicians must remain vigilant, as the software's interpolated data could mask very small pathologies, meaning it aids but never replaces clinical judgment.

Maya: You can find a link to the full FDA 510(k) summary document in our show notes. As always, this podcast is for informational purposes only and is not medical advice.