Company Spotlight: Clarius Ejection Fraction AI - Automating Cardiac Scans at the Bedside

Sam: Welcome to AI in Medicine - Smart Summaries. I am Sam, and today we are looking at a handheld ultrasound AI that actually agreed with the consensus of human experts more than the individual experts agreed with each other, hitting an intraclass correlation coefficient of 0.78.

Maya: I am Maya, and before we begin, a quick reminder that our voices are AI-generated. Today we are summarizing a public document published by the U.S. Food and Drug Administration on March 2, 2026, specifically the FDA 510(k) summary K253593 for the Clarius Ejection Fraction AI by Clarius Mobile Health Corp.

Sam: Exactly. And because this is a Company Spotlight episode, we need to be crystal clear: this is a summary of the FDA's public clearance document, not company marketing material. We are completely independent, and this episode is not sponsored in any way.

Maya: That is right. Our goal is to look at the facts and the real numbers presented to the FDA. The Clarius Ejection Fraction AI is a machine learning algorithm integrated into the Clarius App software. It is designed for use in cardiac ultrasound applications, specifically for the semi-automatic real-time measurement of the left ventricular ejection fraction.

Sam: For anyone who has tried to manually assess a patient's ejection fraction in a busy emergency department or ICU, you know how heavily it relies on deep sonography skills. Having an AI automate that calculation at the bedside is a big deal. How does the document say it actually works?

Maya: The summary explains that during an ultrasound imaging procedure, the user selects a preset, such as Cardiac Basic or Cardiac Advanced, from the Clarius App. Once enabled, the AI places a segmentation mask or landmark markers on the ultrasound image. It specifically identifies the left ventricle in both the End Diastolic and End Systolic phases.

Sam: So it is automatically capturing the heart at its most relaxed and most contracted states to do the math.

Maya: Yes. The document refers to these as the ED and ES frames. Using those frames, the software calculates the ejection fraction of the cardiac images obtained in the Parasternal Long Axis, Parasternal Short Axis, and Apical views, which the text lists as AP4 and AP2.

Sam: Does it just give you a locked-in number, or can the clinician change it?

Maya: The clinician can change it. The user has the option to manually adjust the measurements by moving the caliper crosshairs. The summary explicitly states that the software is an assistive tool intended to inform clinical management and is not intended to replace clinical decision-making. The user retains the ultimate responsibility of ascertaining the measurements based on standard practices and clinical judgment.

Sam: That makes sense. And who is actually cleared to use this? Can anyone with a smartphone use it?

Maya: No. The indications for use specify that the user shall be a healthcare professional trained and qualified in ultrasound. It is also indicated for use in adult patients only.

Sam: Got it. Let's talk about the regulatory pathway. This is a 510(k) clearance, which means the company had to prove their software is substantially equivalent to a legally marketed predicate device.

Maya: Exactly. The FDA document identifies the predicate device as the Caption Interpretation Automated Ejection Fraction Software, manufactured by Caption Health, Inc., which was cleared under 510(k) reference K210747.

Sam: What makes them substantially equivalent in the eyes of the FDA?

Maya: According to the comparison table in the summary, both are Class II devices under the regulation name Medical Image Management and Processing System. They both use artificial intelligence and machine learning for the segmentation of anatomical structures, specifically the heart, and for the measurement of LV ejection fraction. Both operate in professional healthcare settings and are intended for adults.

Sam: Is this a standalone app you can just download from the app store and use on any ultrasound probe?

Maya: No, it is not. The summary clearly states that Clarius Ejection Fraction AI is not a stand-alone software device. It is integrated into the Clarius App software and is only compatible with specific Clarius Ultrasound Scanner system transducers, listing the PA HD3, PAL HD3, and C3 HD3.

Sam: Okay, so it is locked to their hardware ecosystem. Let's get into the most important part: the clinical performance evaluation. How did they actually prove this works?

Maya: They conducted a retrospective analysis of anonymized ultrasound images. The total sample size included in the study was 279 exams.

Sam: 279 exams. Where did that data come from?

Maya: The images were randomly obtained from a multi-center database. The geographic breakdown is actually very detailed. The United States had the most with 72 images. Then Canada with 44, Germany with 22, and interestingly, 21 images were labeled as Unknown.

Sam: Wait, 21 images just had an unknown geographic origin?

Maya: That is what Table 1 shows. Then it lists smaller numbers from many other places. Turkey had 18, the United Kingdom had 10, the Philippines had 9. It goes down to single images from places like South Korea, Jamaica, Israel, Taiwan, and the Vatican.

Sam: It is definitely a globally diverse dataset from an imaging perspective. But what kind of clinical data came with those ultrasound scans? Did they know the patients' ages or underlying cardiac conditions?

Maya: No. The document explicitly states that no clinical or sociodemographic information, such as age, gender, or clinical diagnosis, was available or accessible at any point during the study. This data was fully anonymized prior to their access.

Sam: That feels like a significant limitation. If you are a hospital leader buying an AI for cardiac triage, not knowing if the patient was a perfectly healthy young adult or an elderly patient in severe heart failure means we do not know how the AI performs across specific disease states.

Maya: That is a very fair point. The validation was purely visual. The clinicians who established the ground truth only had access to the ultrasound image for identifying the cardiac anatomy, segmenting the left ventricle, and performing measurements. The summary also notes that the truthing process was not based on any follow-up medical examination.

Sam: Let's talk about the data they left out. Did they include every single image from those exams in the performance testing?

Maya: They did not. The summary outlines specific exclusion criteria. They state that images of inadequate quality were not added to the sample size, specifically noting non-diagnostic images with artifacts obstructing specific anatomy, and images with incomplete anatomy and views.

Sam: So they filtered out the really bad images before comparing the AI to the humans.

Maya: Yes. And another critical detail: in the measurement comparisons, the company excluded the subjects where the Ejection Fraction AI model failed to generate a measurement since there was no value to compare.

Sam: That is exactly why we read these FDA summaries. As a clinician, you need to know that the stellar performance numbers we are about to discuss apply to decent quality images where the AI actually successfully generated a result. It does not tell us how often the AI completely fails to generate a measurement when a resident is scanning a difficult patient on a busy shift.

Maya: Exactly. So let's look at those performance numbers. The primary objective was to determine whether the AI measurements are non-inferior to those obtained manually by human experts. The significance level was set to 0.025, and the equivalence margin was set at 10%, which they also write as 0.10.

Sam: How did the human experts do against each other? We need a baseline to understand how hard this task is.

Maya: To establish that baseline, they calculated the Intraclass Correlation Coefficient, or ICC. The ICC for Reviewer 1 versus Reviewer 2 was 0.67. For Reviewer 1 versus Reviewer 3, it was 0.64. And for Reviewer 2 versus Reviewer 3, it was 0.54.

Sam: So the human experts were hovering between 0.54 and 0.67 in their agreement with each other. That really highlights how subjective manual ejection fraction measurement can be in practice.

Maya: Exactly. Then they compared the AI to the mean of the human reviewers. The ICC for the AI versus the mean reviewers was 0.78.

Sam: Wow. So the AI's agreement with the consensus of the experts was higher than any two individual experts agreeing with each other. 0.78 versus a high of 0.67.

Maya: Yes. The summary concludes that the clinical performance has been adequately verified and determined to be as reliable and accurate as compared to human clinical experts.

Sam: Did they break down the non-inferiority results by the different ultrasound views?

Maya: They did, using a one-sided t-test. For the Apical view, the p-value was 5.57e-21, with a mean difference of -6.27. For the PSAX view, the p-value was 1.57e-36 with a mean difference of -3.87. And for the PLAX view, the p-value was 1.12e-18 with a mean difference of -5.92.

Sam: And all of those tested against that 10% equivalence margin. So the statistics firmly support the non-inferiority claim.

Maya: Correct. But verification on retrospective data is only one part. They also conducted a clinical validation study to evaluate the design and clinical usage. This was done in a simulated use environment using production equivalent units.

Sam: What were they looking for in that simulated environment?

Maya: They wanted to see if users could actually activate the AI using the curvilinear and phased array scanners, image the cardiac anatomy, perform live segmentation, manually adjust the measurements, change the segmentation mask opacity, and display and save the measurement. The results showed consistent results among all users, meeting the pre-defined acceptance criteria.

Sam: Before we move on, did the FDA summary mention anything about cybersecurity? Because an app running on an iOS or Android device definitely carries security risks.

Maya: It did. The summary notes that cybersecurity and vulnerability analyses were conducted. The company implemented a process for preventing unauthorized access, modifications, misuse or denial of use, or the unauthorized use of information that is stored, accessed or transferred from a medical device to an external recipient.

Sam: That is crucial to look for in these clearances. Now, there is one more major section in this document that tech developers and regulatory teams need to pay attention to: the PCCP, or Predetermined Change Control Plan.

Maya: This is a massive advantage for the company. Under section 515C(b)(1) of the Act, if a company wants to modify an already cleared device, they might need to submit a new 510(k). But if the change is consistent with an established PCCP, a new premarket notification is not required.

Sam: Meaning they can update the AI model without going back through the entire FDA clearance process, as long as they follow the plan outlined in this document. What exactly are they allowed to change?

Maya: The summary outlines several planned modifications. The first is the Modification of data input sources. This allows them to add data from current Clarius scanners and future 510(k) cleared Clarius scanners to the AI model so it can be deployed on more scanners.

Sam: What is the risk there? If you swap the hardware, the image geometry could look totally different to the algorithm.

Maya: The document identifies the risk as Data skewing and concept drift. Their mitigation strategy relies on internal testing to ensure that data skewing and concept drift are mitigated.

Sam: What is the next permitted modification?

Maya: The Modification of training hyperparameters. The summary specifically lists the initial learning rate, width multiplier, dropout rate. The rationale is to improve and optimize the model's performance.

Sam: Modifying hyperparameters can definitely lead to overfitting. Does the document address how they will stop that from happening?

Maya: It does. The risk mitigation explicitly states that Proper regularization techniques and cross-validation and dropout will be employed to mitigate overfitting. They will also conduct internal testing and verification to mitigate unintended biases.

Sam: And the other modifications?

Maya: The third is the Modification of post-processing steps. The fourth is the Modification of masked autoencoder architecture and training. Both share the exact same risk profile regarding overfitting and unintended bias, and use the same mitigation strategies we just discussed.

Sam: It is a really comprehensive look at where bedside cardiac AI is heading. You are taking a highly subjective task, automating it on a portable device, and building in a regulatory pathway to continuously improve the model without red tape.

Maya: Exactly. For hospital buyers and clinical leaders, the key takeaway is that this tool demonstrated non-inferiority to human experts with an impressive ICC of 0.78. But you must remember the limitations of the testing: it was a retrospective study with no clinical information available, and they excluded poor quality images and cases where the AI failed to generate a measurement.

Sam: That is the critical context for the clinicians actually using it on the ward. It is an assistive tool to inform your clinical management, not a replacement for your judgment.

Maya: Well said. That wraps up this Company Spotlight on the Clarius Ejection Fraction AI. As always, you can find a link to the full FDA 510(k) summary in our show notes so you can read the document for yourself.

Sam: And remember, this podcast is for informational purposes only. It is not medical advice, and we are not endorsing this or any other product. Thanks for listening, and we will catch you next time.