21 November 2026 · 15 min

Company Spotlight: Csd Labs eMurmur Heart AI - 90.0% Accuracy in Automated Murmur Detection

CardiologyFDA Clearances

We unpack the public FDA 510(k) summary for eMurmur Heart AI (2.2), a software system designed to analyze patient heart sounds and detect abnormal murmurs. We break down the clinical performance data, the software's transition to supporting multiple digital stethoscopes, and how the company's Predetermined Change Control Plan allows for future hardware updates.

Key points

Source: FDA 510(k) summary K252284: eMurmur Heart AI (Csd Labs) - U.S. Food and Drug Administration, 2025

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Company: Csd Labs · Product: eMurmur Heart AI · FDA 510(k) K252284

Your company here. This podcast is looking for its first sponsors: reach clinicians, health-system leaders and medtech and pharma teams following AI in medicine. Sponsorship options and rates →

Transcript

Maya: Welcome to AI in Medicine, Smart Summaries. I am Maya, and today we are looking at an AI tool that achieved 90.0% accuracy in detecting abnormal heart murmurs from digital stethoscope recordings. The single most interesting finding here is how the company leverages a specific regulatory pathway to seamlessly add new hardware over time.

Sam: I am Sam. Right at the top, a quick reminder that our voices are AI-generated and this episode is a summary of a public document. Specifically, we are unpacking the FDA 510(k) summary for eMurmur Heart AI, version 2.2, published by the U.S. Food and Drug Administration on December 19, 2025.

Maya: This is a Company Spotlight episode, but we are completely independent. This is not a sponsored show, and we are not reading from company marketing material. We are looking strictly at the public 510(k) summary submitted by the sponsor, Csd Labs, which is based in Austria. We are going to be clear about exactly what this document proves.

Sam: And it is important to remember that a 510(k) clearance is based on a determination of substantial equivalence to a predicate device, as this summary itself states. So, let us get into it. What exactly is eMurmur Heart AI version 2.2?

Maya: The document defines it as a decision support system for the automated evaluation of recorded patient heart sounds. It is software that takes a digital recording from a stethoscope and analyzes it. The summary explicitly says it identifies the first and second heart sounds, which are noted as S1 and S2, as well as the presence or absence of abnormal heart murmurs.

Sam: When it says it detects the absence of abnormal murmurs, does it define what falls into that category? Because clinically, a murmur is not always a sign of disease.

Maya: Yes, the document clarifies that the absence of abnormal heart murmurs includes cases of no murmurs, as well as innocent murmurs. So it is specifically looking to flag the abnormal ones.

Sam: And who is the intended patient population for this tool?

Maya: It is intended for use on pediatric and adult patients. The user population is defined as a healthcare provider who is licensed or trained to perform auscultation. But there is a very firm boundary drawn around its use.

Sam: Right, the limitations. What are they?

Maya: The summary states it is not intended as a sole means of diagnosis. It says the heart sound interpretations offered by the software are only significant when considered in conjunction with healthcare provider over-read, and including all other relevant patient data.

Sam: That makes total sense for a decision support tool. It is a co-pilot, not a replacement for the clinician. Now, from a technical perspective, how does the workflow actually operate? If I am a clinician on the ward, where is the AI running?

Maya: It does not run directly on the stethoscope itself. The document explains that recordings are made using any supported digital stethoscope, which connects to a front-end client. Those recorded auscultation data are then transmitted over secure, encrypted connections to the eMurmur backend, which hosts the AI.

Sam: So the heavy lifting is done in the cloud, or at least on a server somewhere, and then the results are sent back to the front-end client for the user to see.

Maya: Exactly. The results are returned to the user to support their decision-making process regarding the potential presence of an abnormal heart murmur. The document also notes it can assist users when investigating potential changes among longitudinally recorded heart sounds.

Sam: Wait, really? So it can help track changes over time for a single patient? That is a fascinating detail for managing chronic structural heart conditions.

Maya: The summary mentions it assists when investigating potential changes among longitudinally recorded heart sounds, yes. But it does not provide specific performance data regarding longitudinal tracking in this summary. It just notes that as part of how the user can utilize the results.

Sam: Got it. Sticking just to the text. I also noticed a distinction in the document between the regulated AI and the rest of the software platform. Can you break that down for the software buyers listening?

Maya: This is a great point for health-system IT leaders. The document describes eMurmur as a multiple function software platform. This platform includes apps and a web portal used to stream, record, display, replay, and store auscultation data. The FDA summary explicitly states that those specific functions are not subject to regulatory oversight.

Sam: So basically, moving audio files around and playing them back does not require FDA clearance.

Maya: Correct. But the functions that are subject to FDA premarket review are the eMurmur Heart AI functions. The document states that the FDA assessed the non-regulated functions only to the extent that they could adversely impact the safety and effectiveness of the regulated AI functions.

Sam: That is a very clean architecture for a digital health product. Keep the medical device software separate from the basic data management software. Speaking of data management, what does the summary say about cybersecurity? Hospital buyers are going to ask about that immediately.

Maya: The summary dedicates a section to cybersecurity. It notes that internet traffic is encrypted, and data stored in the backend database is encrypted. The data is also duplicated to another database in a different datacenter. Crucially, it states that no protected health information is stored on the user's devices.

Sam: Okay, so the audio and the results live on the server, not on the clinician's phone or tablet. That definitely simplifies the security footprint on the ward.

Maya: Yes, and it adds that private networking exposes only the eMurmur backend API to the internet, meaning resources like the AI itself or the database are not directly exposed. Users must authenticate and can only access authorized data.

Sam: Let us talk about the hardware. You mentioned it works with any supported digital stethoscope. But digital stethoscopes from different manufacturers probably capture audio very differently. How does the AI handle that variability?

Maya: That is where a very specific software accessory comes in. The summary describes the eMurmur Transformer. It is a software accessory that enables compatibility with multiple digital stethoscope models. The document says recordings made with supported stethoscopes are processed by this Transformer software before being analyzed by the AI.

Sam: So it is basically a translation layer. It normalizes the audio from different devices so the AI algorithm gets a consistent input.

Maya: Exactly. The text states the Transformer ensures consistent analysis across all supported recording devices. This is actually a major shift from the predicate device.

Sam: Right, let us get into the comparison. The new clearance is for version 2.2. The predicate device they compared it to is version 1.0, which was cleared under a different 510(k). What are the main differences between the two?

Maya: Well, for hardware compatibility, version 1.0 only listed one specific stethoscope, the digital stethoscope Littmann 3200. Version 2.2 lists any supported digital stethoscope model alongside that new eMurmur Transformer accessory we just discussed.

Sam: Moving from a single hardware lock-in to a hardware-agnostic platform is a massive business move. What about clinical features? Did the AI itself learn new tricks between version 1.0 and 2.2?

Maya: It did. The upgrade introduces a new feature that determines the average loudness of individual heart sound components. Specifically, it lists S1, systole, S2, and diastole, as well as their respective loudness ratios. The summary notes this new feature has been validated and has no negative effect on safety and effectiveness.

Sam: That is really interesting. Giving clinicians quantifiable metrics on loudness ratios instead of just relying on subjective human hearing. Okay, let us get to the most important part for clinical validation: the performance data. How did they test version 2.2?

Maya: The summary states the dataset used to validate the predicate device was augmented with new, previously unseen data. They explicitly note this was done to minimize the risk of overestimating device performance. And importantly, this resulting dataset was not used during the training of version 2.2.

Sam: That is a critical detail. Testing an AI on the same data it learned from is a big red flag, so it is good the document confirms the validation set was distinct. What was the makeup of this testing dataset?

Maya: It included heart sound recordings from both pediatric and adult subjects. Exactly 50% of these subjects had a confirmed abnormal murmur, while the remaining 50% were confirmed to have no abnormal murmur.

Sam: An exact 50-50 split. We should point out that this is an enriched dataset for testing purposes. In a real-world primary care clinic, you are not going to see a 50% prevalence of abnormal murmurs. The summary does not provide real-world prospective trial data here.

Maya: That is a very fair point, Sam. The summary only gives us the parameters of this specific validation dataset. To evaluate the AI's answers, they needed a gold standard to compare against. The ground truth classification was determined by three board-certified cardiologists based on a majority consensus.

Sam: Majority consensus. Meaning at least two out of the three cardiologists had to agree on whether the recording showed an abnormal murmur or not. So, what were the final numbers? How did version 2.2 actually perform against those cardiologists?

Maya: During validation, the dataset was analyzed by both the new version 2.2 and the predicate version 1.0. For version 2.2, the sensitivity was 90.0%, with a 95% confidence interval ranging from 78.8% to 95.9%.

Sam: Okay, a sensitivity of 90.0%. That means it correctly identified the abnormal murmurs 90.0% of the time. What about specificity? How well did it rule out the normal ones?

Maya: The specificity was also exactly 90.0%, again with a confidence interval between 78.8% and 95.9%. Overall accuracy landed at 90.0%, with the interval between 82.8% and 94.5%.

Sam: Those are solid numbers, but the confidence intervals are fairly wide, dropping down to 78.8% at the low end. It highlights why this is a decision support tool and not an automated diagnostician. How did those numbers compare to the older version 1.0 on the same data?

Maya: Version 2.2 outperformed the predicate across the board. Version 1.0 had a sensitivity of 86.7%, a specificity of 88.3%, and an accuracy of 87.5%. The document concludes that version 2.2 achieves performance that is non-inferior to the predicate.

Sam: So we have a tool that is more accurate and supports multiple hardware models. But there is another regulatory mechanism in this document that really caught my eye. Can we talk about the Predetermined Change Control Plan, or PCCP?

Maya: Absolutely. This is a very modern regulatory approach. The FDA summary states that they reviewed and cleared a PCCP for this device. Under section 515C of the Act, a new premarket notification is not required for a change to a cleared device if the change is consistent with this established plan.

Sam: So if the company wants to add a brand new stethoscope model next year, they do not have to wait months for a new 510(k) clearance, as long as they follow the exact testing steps outlined in this plan.

Maya: Correct. The document specifies that planned modifications under the PCCP are limited to updating the eMurmur Transformer to support new stethoscope models. It states that adding a new model does not affect already supported models, and the AI itself is not affected by the modification.

Sam: But they still have to prove it works, right? They cannot just flip a switch. What does the summary say about the testing required for a new stethoscope?

Maya: They have to perform very specific validation activities. The plan requires collecting recordings from patients at multiple locations using the new stethoscope model. The sample size must be determined with a power of 80% and a confidence level of 95%.

Sam: And do they use the same standard for the ground truth? The cardiologists?

Maya: Yes, the document states the ground truth classification, meaning whether there is an abnormal murmur present or the patient is healthy, must be determined by three board-certified cardiologists through majority consensus. Validation is only successful if it demonstrates non-inferior AI performance compared to an already supported stethoscope model.

Sam: And how do they prove non-inferiority statistically in this plan?

Maya: The summary notes non-inferiority is established using confidence intervals for unpaired data with continuity correction for the true difference in sensitivity or specificity. If the validation is not successful, the modification is not implemented, and the stethoscope model will not be supported.

Sam: That is a fantastic look under the hood at how the FDA is managing the rapid pace of digital health innovation. It gives the company agility, but hardcodes the statistical rigor required to keep patients safe. How do users find out when a new stethoscope is added?

Maya: The document says users are informed through updated labeling. Specifically, a table listing all supported stethoscopes is provided and updated whenever a new model is added. This table is accessible through the User Manual.

Sam: Before we wrap up, I want to touch on the standard regulatory boilerplate at the beginning of the letter. It mentions this is a Class II device under regulation number 21 CFR 870.1875. It also lists several general controls.

Maya: Yes, it reminds the sponsor that clearance does not mean exemption from the Quality System regulation, found in 21 CFR Part 820. This includes strict rules for design controls, handling nonconforming product, and corrective and preventive action. They must also comply with the Unique Device Identification System rule.

Sam: Right, the UDI rule requires a unique identifier on the label unless an exception applies. And the FDA letter points out that deviations from the PCCP that result in major changes could significantly affect safety or effectiveness, meaning a new submission would be required. Failure to submit would constitute adulteration and misbranding under sections 501 and 502 of the Act.

Maya: Exactly. The regulatory guardrails remain very firmly in place. To recap, the eMurmur Heart AI version 2.2 is cleared to assist healthcare providers by identifying S1, S2, and abnormal murmurs, while also calculating loudness metrics. It demonstrated 90.0% accuracy in its validation testing and utilizes a smart software architecture to support multiple digital stethoscopes safely.

Sam: A great example of AI moving from theoretical potential into structured, regulated clinical workflows. Remember, you can find a link to the full FDA summary in the show notes. And as always, this podcast is for informational purposes only and is not medical advice.

Maya: Thanks for joining us on AI in Medicine, Smart Summaries. We will catch you next time.