FDA Clears AI for Heart Failure: Unpacking the AiTiALVSD 510(k)

Sam: Imagine running a standard, routine 12-lead ECG, and an algorithm tells you with 83.7 sensitivity that the patient has a weakened heart muscle.

Maya: That is exactly what we are looking at today. We are breaking down the official U.S. Food and Drug Administration 510(k) clearance document for a device called AiTiALVSD, published September 30, 2026.

Sam: And before we get into the weeds of this clearance, a quick reminder that our voices are AI-generated, and this episode is a summary of a publicly available document.

Maya: The company behind this is Medical AI Co., Ltd., based in Seoul, Republic Of Korea. They submitted this device, and the FDA has determined it is substantially equivalent to legally marketed predicate devices, allowing them to market it subject to general controls.

Sam: So let us start with the basics. What exactly is AiTiALVSD designed to do?

Maya: It is an electrocardiogram analysis software. It uses artificial intelligence algorithms to support the diagnosis of Left Ventricular Systolic Dysfunction, or LVSD.

Sam: And the document is very specific about how it defines that dysfunction, right?

Maya: Exactly. The FDA document defines LVSD here as a Left Ventricular Ejection Fraction less than or equal to 40%.

Sam: Which is a critical threshold in cardiology. But they are not using an echocardiogram to find this, which is the traditional way. They are using 12-lead ECG data.

Maya: Right. The user uploads the ECG data in XML format, or it can be automatically transmitted from an ECG medical device. The software then analyzes that data and gives a binary output. It just says High or Low.

Sam: High or Low. So it is not giving you a specific ejection fraction percentage. It is just flagging risk.

Maya: Exactly. The output indicates a high or low likelihood of an LVEF less than or equal to 40% by echocardiography. And the FDA classifies this under the regulation name Cardiovascular Machine Learning-Based Notification Software. It is a Class II device.

Sam: I want to dig into who this is actually for. The indications for use section says it is for adults at risk for heart failure. Who falls into that bucket according to this clearance?

Maya: It is a pretty broad list. The target population includes, but is not limited to, individuals with coronary artery disease, diabetes mellitus, cardiomyopathies, hypertension, and obesity.

Sam: That covers a massive portion of adults walking into any primary care clinic or emergency room.

Maya: It does. And the list continues. It also includes people with a history of myocardial infarction, aortic stenosis, atrial fibrillation, exposure to cardiotoxic pharmaceutical therapies, cardiac hypertrophy evident in chest X-rays, and postpartum women.

Sam: Postpartum women is an interesting inclusion. But there is also a specific exclusion mentioned right after that list.

Maya: Yes. The document explicitly states that AiTiALVSD should not be used on ECGs with a paced rhythm.

Sam: Because a pacemaker artificially alters the electrical signals of the heart, which would confuse an algorithm trained on natural rhythms.

Maya: Exactly. And that brings us to the warnings section, which is critical for clinicians and health systems to understand before buying or deploying this tool.

Sam: The FDA is very firm here. I am looking at Warning A. It says users should not rely solely on the device output to determine patient management or the need for follow-up.

Maya: And Warning B reinforces that. It says the output should be interpreted only in conjunction with a clinician's professional judgment and the patient's full clinical evaluation.

Sam: They even explicitly state that AiTiALVSD results are not a diagnosis and must not be used as a substitute for echocardiography or other clinical diagnostic tests.

Maya: That is the reality of Class II machine learning notification software. It is a flag, a support tool. It is not replacing the gold standard test, which is the echo. It is just helping you decide who might need that echo sooner rather than later.

Sam: Let us talk about how this AI was actually built. Where did the data come from?

Maya: The training dataset is fascinating. It came from 16 hospitals in Korea. They used paired ECGs and TTE, which is echocardiography, to ensure accurate labeling.

Sam: And the sheer volume of data is impressive. The document notes the training dataset comprised approximately 148,624 patients.

Maya: And those patients contributed 400,339 ECGs. They used this data to train a deep learning neural network with transformer-based self-attention mechanisms via supervised learning.

Sam: Transformer-based self-attention. That is the exact wording in the text. It means the model is looking at the entire ECG waveform and learning which specific parts of that electrical signal are most strongly associated with a low ejection fraction.

Maya: Exactly. And the text clearly states the model was trained to identify patients with LVSD based solely on ECG data.

Sam: So that is the training side. 148,624 patients in Korea. But to get FDA clearance, they had to prove it works in a US population, right?

Maya: Yes. The performance of the software was evaluated in a US retrospective study. The document emphasizes that no data present in the training dataset was used in the performance study.

Sam: Let us look at the demographics of that US performance study. Table 1 breaks it all down. How many patients were involved?

Maya: The performance study included 22,101 patients. And that meant 22,101 ECG samples, so one per patient.

Sam: The mean age was 65, with a standard deviation of 16. The minimum age was 18, and the maximum was 106.

Maya: It is a broad age range, and they actually break down the age brackets in the table. There were 1,942 patients under age 40, which is 8.8%.

Sam: Then 1,844 in the 40 to 49 range. But you can see it really scales up with age. 3,173 in their fifties.

Maya: And the two largest buckets were the sixties and seventies. There were 5,308 patients aged 60 to 69, and 5,384 aged 70 to 79. Both represent 24% of the study.

Sam: Even the 80 and older group is substantial. 4,450 patients, or 20% of the study. What about the sex breakdown?

Maya: It was a perfect split in the US study. 10,991 female, which is 50%, and 11,110 male, which is also 50%.

Sam: That is a great split. Interestingly, the document notes that the Korean training dataset was a bit different. It was 45% female and 55% male.

Maya: That happens often when moving from training to validation datasets. Let us talk about the race and ethnicity demographics in the US study, because the FDA highlights something really important here.

Sam: Right. Looking at the race breakdown, 16,507 patients were White. That is 78%.

Maya: The next largest group was Black or African American at 2,026 patients, which is 9.6%. Then Asian at 935 patients, or 4.4%.

Sam: There were also 422 American Indian or Alaska Native patients, making up 2.0%, and 44 Native Hawaiian or Other Pacific Islander patients, which is 0.2%. Plus 1,180 listed as Other or Unknown, and 987 with missing information.

Maya: Quick note before we carry on. This spot is open for a sponsor. If your company builds or sells AI for healthcare and wants to reach the clinicians, health-system leaders and industry teams who listen to this show, the link to our sponsorship page is in the show notes.

Sam: And now, back to the document.

Maya: For ethnicity, 2,478 patients were Hispanic or Latino, which is 11%, compared to 19,247 who were Not Hispanic or Latino, which is 89%. 376 were unknown or missing.

Sam: The FDA explicitly calls out a limitation regarding that racial breakdown in a later paragraph. What did they find?

Maya: The document states that the clinical validation study included a limited number of Black or African American subjects. Again, they note that number as 2,026, or 9.6% of subjects with recorded race.

Sam: And the critical sentence is this: In this subgroup, the specificity was lower than in the overall study population.

Maya: That is a vital piece of information for any health system leader. If you serve a large Black or African American population, you need to know that the algorithm's specificity—its ability to correctly identify people without the condition—was lower in this group.

Sam: Meaning there might be a higher rate of false positives in that specific population, leading to unnecessary follow-up echoes. The document mentions the study was not powered to establish performance in this population separately.

Maya: Exactly. Let us look at the overall clinical characteristics of the 22,101 patients in the US study. They were a sick population.

Sam: Very. Hypertension was the most common condition at 65%, which is 14,337 patients. Coronary Artery Disease was 40%, or 8,939 patients.

Maya: Diabetes Mellitus was at 30%, which is 6,602 patients. Congestive Heart Failure was 29%, or 6,309 patients.

Sam: They also tracked Atrial fibrillation at 26%, Chronic Kidney Disease at 22%, and Myocardial Infarction at 17%. The mean BMI was 29.

Maya: So this software was tested on a real-world, complex patient population. Now, let us get into the actual performance metrics. How well did it work?

Sam: The results for the overall population are in Table 2. The Sensitivity was 83.7, and the Specificity was 81.0.

Maya: To put that in context, a sensitivity of 83.7 indicates its ability to correctly flag people who actually have an ejection fraction less than or equal to 40%.

Sam: And the specificity of 81.0 reflects its ability to correctly identify those who do not have the condition.

Maya: But in clinical practice, we really care about predictive values. The Positive Predictive Value, or PPV, was 34.5, and the Negative Predictive Value, or NPV, was 97.7.

Sam: Those numbers are based on the observed prevalence of LVSD in the US clinical study, which the document notes was 10.5%.

Maya: Yes, and the FDA included a really helpful chart showing how PPV and NPV change based on different prevalences in a given population. This is exactly what hospital administrators need to look at.

Sam: Let us walk through that. If a clinic has a low LVSD prevalence, say 6%, the PPV is 0.219.

Maya: That means if the AI says High, there is a lower chance the patient actually has LVSD. But the NPV at 6% prevalence is 0.987.

Sam: So if the AI says Low, that 0.987 Negative Predictive Value means the clinician can be highly confident in ruling it out.

Maya: Exactly. Let us look at the middle of the chart. At an 8% prevalence, PPV is 0.277 and NPV is 0.983. At 10%, PPV is 0.328 and NPV is 0.978.

Sam: And the numbers keep shifting as prevalence goes up. At 14%, PPV rises to 0.417 and NPV drops slightly to 0.968. At 16%, PPV is 0.456 and NPV is 0.963.

Maya: Finally, if you deploy this in a high-risk cardiology clinic where the prevalence is 20%, the PPV reaches 0.524, and the NPV is 0.952.

Sam: So regardless of the prevalence, from 6% all the way up to 20%, the negative predictive value stays incredibly high, with the lowest being 0.952.

Maya: Which makes this tool potentially very useful for ruling out LVSD. If the software says Low, a clinician can be highly confident, though again, not certain, that the patient's ejection fraction is above 40%.

Sam: Let us talk about the regulatory pathway here. This was a 510(k) submission, which means Medical AI Co., Ltd. had to prove their device was substantially equivalent to a predicate device already on the market.

Maya: Right. The predicate device they used is the Anumana, Inc. Low Ejection Fraction AI-ECG Algorithm. Its submission number is K250652.

Sam: The document has a Substantial Equivalence Comparison table, Table 2, comparing AiTiALVSD directly with the Anumana device. Both are Rx only. Both are for adults.

Maya: And both use a locked machine-learning based model. That means the algorithm does not change or learn on its own once it is deployed in the hospital.

Sam: The inputs are the same too. Both require a 12-lead ECG waveform in digital format. Specifically, they need compatible 12-Lead diagnostic ECG machines with 500Hz digital output.

Maya: That 500Hz digital output is important for hospital IT departments to note. You cannot just use any old ECG machine; it needs to meet that hardware requirement.

Sam: How do the clinical performance numbers compare between the two?

Maya: They are remarkably close, which the FDA notes is why they are substantially equivalent. We mentioned AiTiALVSD has a sensitivity of 83.7. The Anumana predicate device has a sensitivity of 84.5%.

Sam: AiTiALVSD specificity is 81.0, and Anumana is 83.6%.

Maya: For the predictive values, AiTiALVSD's PPV is 34.5, while Anumana's is 30.5%. And for NPV, AiTiALVSD is at 97.7, compared to Anumana's 98.4%.

Sam: So they are truly neck and neck. The FDA writes that differences in the indications for use statement do not result in a new intended use, and any differences in technological characteristics do not raise different questions of safety or effectiveness.

Maya: There is a slight difference in how the outputs are displayed. AiTiALVSD provides an output of Low-risk group for Left Ventricular Systolic Dysfunction, High-risk group for Left Ventricular Systolic Dysfunction, or Error.

Sam: While the Anumana device outputs Low LVEF Detected, Low LVEF Not Detected, or Error. But functionally, they are doing the exact same thing.

Maya: One more regulatory detail worth noting. The letter mentions that the device is subject to the Quality Management System Regulation, or QMSR. This includes ISO 13485 clauses for design controls, nonconforming product, corrective action, and preventative action.

Sam: It also reminds the manufacturer that device labeling must be truthful and not misleading, and they must comply with medical device reporting for adverse events.

Maya: This document gives us a perfect snapshot of where cardiovascular AI stands right now. We are seeing models trained on massive datasets—like those 400,339 ECGs in Korea—being validated on diverse, real-world US populations of 22,101 patients.

Sam: And the FDA is clearing them with very clear boundaries. They are Class II tools, they provide notification, not diagnosis, and they must be used alongside professional judgment.

Maya: And perhaps most importantly, the FDA is paying attention to demographic disparities, explicitly noting the lower specificity in the 2,026 Black or African American subjects in the validation study. That transparency is crucial for safe clinical adoption.

Sam: Absolutely. That wraps up our deep dive into the 510(k) clearance for AiTiALVSD. We have linked the full FDA document in the show notes so you can review the data tables for yourself.

Maya: As always, this podcast is for informational purposes only and is not medical advice. Thanks for joining us on AI in Medicine - Smart Summaries.