Company Spotlight: Eko Health EFAST - Transformers hit the digital stethoscope

Sam: Welcome to AI in Medicine - Smart Summaries. Before we get into today's deep dive, a quick reminder that our voices are AI-generated.

Maya: And today we are examining a public document published by the U.S. Food and Drug Administration on August 12, 2025. It is the FDA 510(k) summary for a product called the Eko Foundation Analysis Software with Transformers, or EFAST.

Sam: The hook here is the word 'Transformers' right in the title. We are seeing foundation models officially cleared by the FDA to analyze heart sounds and ECGs from a digital stethoscope. Just to be clear, this is a summary of the FDA's public clearance document, not an advert or company PR material.

Maya: Exactly. This is a Company Spotlight episode where we only look at what the FDA summary actually says. And the clearance here is based on 'substantial equivalence', meaning the FDA determined EFAST is as safe and effective as legally marketed predicate devices.

Sam: Let's start with the basics. What exactly is the EFAST software and who is it for?

Maya: The document defines it as a cloud-based Software as a Medical Device intended to provide clinical decision support to healthcare professionals. It evaluates a patient's heart sounds, known as a phonocardiogram, and their electrocardiograms.

Sam: And the input for this is a 15-second recording captured using Eko Digital Stethoscopes. The summary specifically names the Eko CORE and CORE 500.

Maya: Right. Those devices save the recording in a dot-WAV file format, and then it gets sent to the Eko Cloud.

Sam: Which the document notes uses the Amazon Web Services S3 for storage. I love that they include that detail. IT leaders at health systems always want to know exactly where the data is sitting.

Maya: From there, an API Request is sent to the EFAST API. The deep neural networks analyze the recordings, and then the software returns a JSON format data structure with the algorithm results.

Sam: Wait, really? Just a JSON file? So there isn't a screen on the stethoscope itself that lights up with the answer?

Maya: No. The document explicitly states under the interface section that there is 'No primary display'. The JSON is saved back to the Eko Cloud and can be displayed on EFAST API integrated target software or user-facing apps.

Sam: That is a huge shift. The stethoscope is acting as a data acquisition node, and the actual medical device is an API in the cloud analyzing the acoustic and electrical signals.

Maya: Exactly. And the software consists of five algorithm components. First is a Heart Sound Signal Quality Classifier, which evaluates if the audio is actually good enough for clinical analysis.

Sam: A necessary gatekeeper. If the signal is bad, you don't want a neural network trying to diagnose it.

Maya: Then there is the Structural Murmur Classifier, which determines if a good-quality recording contains a structural murmur. There is also a component for Heart Sound Timing and Murmur Timing, an ECG Rhythm Classifier, and a Heart Rate Calculation algorithm.

Sam: Let's dig into the indications for use. Who are these specific classifications for?

Maya: The murmur detection is intended for both pediatric and adult patients. But the detection of atrial fibrillation from the ECG is only intended for adults, meaning patients over 18 years old.

Sam: And there is a very clear warning that this software is not to be used as a sole means of diagnosis.

Maya: Right. The interpretations are designed to support and not replace the healthcare professional's clinical judgment. It also specifies that the software does not identify other arrhythmias outside of atrial fibrillation and normal sinus rhythm.

Sam: Let's talk about how they proved this works to the FDA. The testing for the Structural Murmur Classification was a retrospective study of de-identified data.

Maya: Yes, they used 2,460 unique heart sound recordings from 615 unique subjects across three sites. Two in the US and one in Canada.

Sam: But they didn't use all of them. The document says that for 259 of the 2,460 heart sound recordings, the cardiologists annotating the data could not come to agreement.

Maya: So those 259 recordings were simply excluded from analysis.

Sam: Which is fascinating. The ground truth was paired with a gold standard echocardiogram, but if the human experts couldn't agree on what they were hearing, they just tossed the file. The document doesn't give a medical explanation for why those specific files were so contentious.

Maya: It doesn't. But of the recordings they kept, which experts identified as good signal, 49.5% had a confirmed structural murmur and 50.5% had no structural murmur.

Sam: So almost an even split in the test data between confirmed murmurs and no structural murmur. What was the performance?

Maya: Overall sensitivity for structural murmur classification was 83.4%. Overall specificity was 86.0%.

Sam: How does that compare to the older software? Because this is a 510(k) based on substantial equivalence.

Maya: The primary predicate device, the Eko Murmur Analysis Software, had a sensitivity of 85.8% and a specificity of 84.5%. So the new EFAST model is very comparable in its overall numbers.

Sam: But the subgroup analysis is where a clinician really needs to pay attention, especially regarding pediatric patients.

Maya: Absolutely. When they broke the results down by age group, adults 18 and older saw a specificity of 93.9%. Adults 60 and older had a specificity of 90.9%.

Sam: But for pediatric patients, under 18, the specificity dropped significantly to 76.0%.

Maya: The summary provides a direct reason for this. It states this is 'likely due to confusion of structural murmurs with some innocent murmurs on pediatric patients.'

Sam: Which means if you are screening a child, you have a much higher chance of a false positive. And that plays directly into the Positive Predictive Value table they included.

Maya: Yes, they provide a table adjusting the Positive Predictive Values based on an assumed prevalence of 0.25 for Structural Heart Disease.

Sam: Meaning they calculate how the tool performs if you assume a prevalence of 0.25 in the tested population.

Maya: Exactly. At an assumed prevalence of 0.25, the adjusted PPV for a 'High Likelihood' prediction across all ages is 76.0%.

Sam: And for adults 18 and older, that adjusted PPV for a 'High Likelihood' prediction is 88.9%. That's fairly strong.

Maya: But look at the pediatric row. For children, the adjusted PPV for a 'High Likelihood' prediction drops to 65.4%.

Sam: And if the software just gives a 'Moderate Likelihood' prediction for a pediatric patient, the adjusted PPV is only 23.2%.

Maya: That is a vital number for any clinician on the ward. A moderate likelihood prediction on a pediatric patient with an assumed 0.25 prevalence means the positive predictive value is exactly 23.2%.

Sam: We also need to look at the demographics of the test data. The document gives us the exact breakdown.

Maya: For the structural murmur study, race was documented for 458 patients. Of that subset, 412 were White, which is 90.0%.

Sam: Black or African American patients made up 22 individuals, or 4.8%. Other was 15, or 3.3%. Asian was 6, or 1.3%.

Maya: Native Hawaiian or Other Pacific Islander was 2 patients, which is listed as <1%, and American Indian or Alaska Native was 1 patient, also <1%.

Sam: So the dataset is overwhelmingly White. The FDA summary says this is 'therefore representative of the intended use population.'

Maya: They did run a subgroup assessment and stated 'No significant differences were observed' across race or ethnicity. But health system buyers should be aware that the sample sizes for non-White patients were extremely small.

Sam: Let's move to the second major feature: the ECG Rhythm Classification for detecting atrial fibrillation.

Maya: This was also a retrospective study of de-identified data. They used 1,256 ECG recordings from 314 unique subjects at two US sites.

Sam: Collected using the Eko CORE 500. The median age here was 67.5.

Maya: And just like the murmur study, there were exclusions. For 97 of the 1,256 recordings, the multiple experts—which included electrophysiologists and certified cardiographic technicians—did not come to an agreement.

Sam: So those 97 recordings were thrown out. Of the remaining good signal recordings, 23.5% had confirmed atrial fibrillation and 76.5% did not.

Maya: This is where we see the biggest performance leap. EFAST achieved an overall sensitivity of 94.7% and a specificity of 94.1% for detecting atrial fibrillation.

Sam: And the reference device they compared it to, the Eko Analysis Software, had a sensitivity of 73.1%.

Maya: Going from a sensitivity of 73.1% up to 94.7% is a massive improvement. The older device did have a slightly higher specificity at 97.7%, compared to EFAST at 94.1%.

Sam: But for a screening tool, catching significantly more true positives is usually exactly what you want from an underlying neural network upgrade.

Maya: On the demographics for the ECG study, it was heavily skewed again: 272 patients were White, which is 86.6%. 19 patients were Black or African American, which is 6.1%.

Sam: Other was 15 patients at 2.4%, Asian was 3 patients at 1.0%, Native Hawaiian or Other Pacific Islander was 2 at <1%, and American Indian or Alaska Native was 1 at <1%. Race information was unavailable for 2 patients.

Maya: They noted no significant differences in performance across biological sex either. The ECG sample was 63.1% male and 36.9% female.

Sam: Whereas the murmur dataset was slightly different, sitting at 57.7% male and 42.3% female.

Maya: There are two smaller components tested too. EFAST calculates heart rate, and the mean absolute percentage error was found to be less than 5%.

Sam: It also determines the timing of the S1 and S2 heart sounds. Based on 96 recordings annotated by experts, S1 detection sensitivity was 98.58% and S2 sensitivity was 94.81%.

Maya: And under the regulatory information, we see this device has a Regulatory Class of Class II, and is categorized under product codes DQD, DQC, DPS, and MWI.

Sam: The document doesn't actually explain what those specific product codes stand for, but they are the FDA's internal buckets for classifying this specific type of technology.

Maya: Right. Ultimately, the FDA concluded that EFAST is substantially equivalent, as safe and effective as the predicate device, and raises no new safety or effectiveness issues.

Sam: So, what's the memorable takeaway here for a health system leader or a medtech developer?

Maya: It is that the digital stethoscope has officially been unbundled. The hardware captures the sound, but the diagnostic power lives in a cloud-based transformer model returning a JSON file via an API. It is built to disappear into the software ecosystem a hospital already uses.

Sam: But the trade-off is knowing the limitations of the data it was trained and tested on. The testing was entirely retrospective, they excluded files where human experts disagreed, and the specificity drops to 76.0% for pediatric murmurs.

Maya: Which is exactly why the FDA clears it as a decision support tool, not a replacement for a doctor's ears and judgment.

Sam: To recap, today we unpacked the FDA 510(k) summary for Eko Health's EFAST software, looking at its cloud-based architecture and its performance numbers for murmur and AFib detection.

Maya: You can find a link to the full public source document in the show notes. And remember, this podcast is for informational purposes only and is not medical advice.