Company Spotlight: Pathwai EchoNext - Finding Structural Heart Disease on a Routine ECG

Maya: Welcome to AI in Medicine - Smart Summaries. I am Maya, my co-host is Sam, our voices are AI-generated, and today we are summarizing a public document that shows how a routine ECG might soon flag hidden structural heart disease.

Sam: Exactly. We are looking at a 510(k) summary published by the U.S. Food and Drug Administration on June 8, 2026, clearing a device called EchoNext, made by a company called Pathwai.

Maya: And just to be completely clear up front, this is a company spotlight episode based entirely on the FDA public clearance document, not on any marketing or PR material from the company itself.

Sam: So let us get right into what EchoNext is. It is software-as-a-medical-device, and it analyzes standard resting 12-lead electrocardiograms.

Maya: Right, the 12-lead ECG is something we do constantly in the clinic. It is cheap and fast, and we normally use it to look for arrhythmias or acute events. But EchoNext is doing something different. It is looking for structural heart disease.

Sam: The summary says it processes the digital 12-lead ECG waveform data using a machine-learning algorithm based on deep neural network techniques.

Maya: And it combines that waveform data with basic demographic inputs. The scientific basis, according to the document, is the established relationship between subtle electrical patterns in the ECG signals and underlying structural abnormalities of the heart chambers, walls, or valves.

Sam: That makes sense. The electrical signal has to travel through the heart structure, so if the structure is changed by disease, the electrical pattern changes, even if it is too subtle for the human eye to catch.

Maya: Exactly. The software looks for those subtle signs, and if it finds them, it generates a notification indicating whether the ECG demonstrates patterns associated with a suspicion of structural heart disease.

Sam: The document abbreviates structural heart disease as SHD. And the FDA summary notes that when SHD is suspected, the device may also provide informational indicators for specific affected cardiac structures. It outputs a primary indicator for SHD, and then it has six specific sub-labels.

Maya: Let us walk through those sub-labels, because clinically, structural heart disease is a huge umbrella. The first sub-label is LVEF, or Left Ventricular Ejection Fraction, specifically when it drops below 45%.

Sam: Next is LVH, Left Ventricular Hypertrophy, specifically severe LVH. Then there is AS, Aortic Stenosis, at moderate or greater severity.

Maya: It also flags Valvular Regurgitation, again at moderate or greater severity. The document specifies that this includes Aortic Regurgitation, Mitral Regurgitation, Pulmonary Regurgitation, and Tricuspid Regurgitation.

Sam: The fifth sub-label is RVEF, Right Ventricular Ejection Fraction, looking for moderate or greater dysfunction.

Maya: And finally, PH, or Pulmonary Hypertension, defined as a pulmonary artery systolic pressure greater than or equal to 45mmHg or a tricuspid regurgitation max velocity greater than or equal to 3.2 meters per second.

Sam: So, the algorithm takes a 12-lead ECG, adds demographic data, and outputs a binary notification for SHD, plus potentially these six sub-labels.

Maya: Correct. And the intended use is for patients 18 years of age or older. It is intended to be used in parallel to the standard of care.

Sam: The document has a massive section on the clinical performance data. They validated this on a very large dataset.

Maya: They did. The analysis uses a dataset of 102,219 unique consecutive patients, with one ECG per patient.

Sam: The summary describes this as an "external, prospectively enrolled cohort" across four independent clinical sites.

Maya: Wait, let me look closely at the wording there. In one paragraph it says EchoNext was validated based on a "retrospective, multicenter evaluation". But then in the performance analysis section, it says it was validated in an "external, prospectively enrolled cohort" of 102,219 patients. That is an interesting phrasing discrepancy in the FDA summary itself, but either way, the performance testing was conducted on 102,219 patients.

Sam: And the data comes from four geographically distinct health systems in the United States and Canada. The summary emphasizes that these sites span academic and community hospitals, covering urban, suburban, and rural settings.

Maya: They aimed to maximize geographic, demographic, and clinical diversity. The median patient age was 66.3 years, and 52.0% were male.

Sam: The document also gives us the race and ethnicity breakdown. The dataset was 41.4% White, 17.4% Asian, 10.8% Hispanic, and 6.1% Black.

Maya: Out of those 102,219 patients, the prevalence of structural heart disease was 34.8%. That means 35,548 patients in the dataset actually had SHD.

Sam: How did they know for sure if a patient had SHD? What was the ground truth?

Maya: The clinical ground truth was established using transthoracic echocardiography. The document calls this the "diagnostic reference standard" for structural heart disease. Board-certified cardiologists interpreted the echocardiograms using FDA-cleared systems, and they were blinded to the ECG findings.

Sam: So they took finalized echo reports, extracted structured measurements, and used guideline-based thresholds to assign the SHD label and the sub-labels. And the rule was that patients had to have a clinically ordered 12-lead ECG and a transthoracic echocardiogram within 30 days.

Maya: That is a robust way to build a dataset. You have the cheap, fast test, the ECG, and you compare it against the diagnostic imaging, the echo, done within 30 days.

Sam: Let us look at the actual performance numbers. How well did EchoNext do at finding structural heart disease?

Maya: For the composite structural heart disease output, EchoNext achieved a sensitivity of 73.7%.

Sam: And what was the confidence interval on that?

Maya: The 95% confidence interval was 73.2 to 74.1%.

Sam: And the specificity?

Maya: Specificity was 75.2%, with a confidence interval of 74.9 to 75.5%.

Sam: Okay, so a sensitivity of 73.7% and a specificity of 75.2%. What does that actually mean for a clinician looking at this software's output?

Maya: It means that if a patient truly has structural heart disease, EchoNext will flag them about 73.7% of the time in a population like this. And if they do not have it, it will correctly stay silent 75.2% of the time. It is not perfect, but remember, the ECG is not traditionally used to diagnose these structural issues at all.

Sam: The positive predictive value, or PPV, for SHD in this dataset was 61.6%. And the negative predictive value, the NPV, was 84.6%.

Maya: The sub-labels also have their own specific performance metrics. For example, for Left Ventricular Systolic Dysfunction, the sensitivity was 80.1% and the specificity was 80.4%.

Sam: The NPV for that left ventricular sub-label was very high, 97.0%. But the PPV was only 34.8%.

Maya: The prevalence for Left Ventricular Systolic Dysfunction in the dataset was 11.3%. Because the prevalence is relatively low compared to the overall SHD bucket, the positive predictive value drops. That means if the device flags left ventricular systolic dysfunction, there is only a 34.8% chance the patient actually has it based on this dataset's prevalence.

Sam: The summary actually addresses how prevalence changes the predictive values. It provides estimates for how PPV and NPV would change at different prevalences for some sub-labels.

Maya: Yes, for valvular regurgitation, they estimate that at a prevalence of 8.2%, the PPV would be 24.4% and the NPV would be 95.7%. If the prevalence drops to 5.0%, the PPV drops to 16.0% and the NPV goes up to 97.4%.

Sam: For Pulmonary Hypertension, it estimates that at a prevalence of 4.3%, the PPV would be 13.2% and the NPV 98.0%.

Maya: I appreciate that the FDA summary includes those estimates, because in the real world, the prevalence of these diseases changes drastically depending on whether you are in a specialized cardiology clinic or a general primary care office.

Sam: What about the other sub-labels? How did it do on Right Ventricular Systolic Dysfunction?

Maya: For Right Ventricular Systolic Dysfunction, sensitivity was 76.8% and specificity was 82.5%. The prevalence was only 3.0%, so the PPV was 12.1% and the NPV was 99.2%.

Sam: And Aortic Stenosis?

Maya: Aortic Stenosis had a lower sensitivity at 54.9%, but a specificity of 80.1%.

Sam: Valvular Regurgitation?

Maya: Sensitivity of 58.2%, specificity of 83.9%.

Sam: And finally, Left Ventricular Hypertrophy.

Maya: Sensitivity was 48.2%, and specificity was 81.3%.

Sam: So across the board, the specificities for the sub-labels range from 80.1% to 83.9%, while sensitivities range from 48.2% for LVH up to 80.1% for Left Ventricular Systolic Dysfunction.

Maya: Exactly. The FDA summary notes that pre-specified acceptance criteria were defined separately for sensitivity and specificity for each output prior to testing, and EchoNext met all pre-specified acceptance criteria across the composite SHD output and all six sub-labels.

Sam: They also did sub-group analyses. They stratified device performance by age, sex, ECG manufacturer, comorbidity, race and ethnicity, and baseline ECG to index TTE. And in all circumstances, EchoNext met its acceptance criteria.

Maya: We also have to talk about the predicate device. To get a 510(k) clearance, a company has to show their device is substantially equivalent to something already legally marketed.

Sam: Right. The predicate device listed here is the Viz HCM Module, under the regulation number DEN230003.

Maya: The summary provides a side-by-side comparison. Both are software-as-a-medical-device under the same regulation name, "Cardiovascular Machine Learning-Based Notification Software". Both take a 12-lead ECG as input.

Sam: But there are differences. Viz HCM is specifically for detecting signs associated with hypertrophic cardiomyopathy, or HCM. It outputs a binary notification for an HCM flag. EchoNext looks for a broader composite of Structural Heart Disease, with those six sub-labels.

Maya: Another difference noted in the comparison table is that Viz HCM has a built-in view, whereas EchoNext does not.

Sam: The FDA determined that despite these differences, both devices have the same intended use, similar principles of operations, and technological characteristics. The differences do not raise different questions of safety and effectiveness.

Maya: Now, let us talk about the limitations, because the FDA clearance is very specific about how this should not be used.

Sam: First off, the indications for use explicitly state that EchoNext is not intended for use on patients with implanted pacemakers. The document does not explain why, it just states it as a rule.

Maya: The document also states firmly that EchoNext is limited to analysis of ECG data and should not be used in lieu of full patient evaluation or relied upon to make or confirm a diagnosis.

Sam: It identifies patients for further SHD follow-up and is not intended to replace the current standard of care methods.

Maya: And crucially, the results of the device are not intended to rule-out SHD follow-up. This is an important distinction. A negative result from EchoNext does not mean you can definitively say the patient does not have structural heart disease. If clinical suspicion is high, you still follow the standard of care.

Sam: What other testing did they do before getting cleared?

Maya: They did Software Verification and Validation. The document states that 101 out of 101 test cases passed.

Sam: They also did Cybersecurity Testing, including Static and Dynamic Code Analysis, Fuzz Testing, and Penetration testing, and found no vulnerabilities.

Maya: And they did Human Factors testing with 15 participants representative of the intended user population. They found zero use errors or close calls.

Sam: So looking at the big picture, we have a machine learning tool that processes a routine 12-lead ECG to flag possible structural heart disease, cleared based on a dataset of 102,219 patients.

Maya: The scale of that dataset is significant. Having 102,219 patients with matched ECGs and echocardiograms within 30 days provides a robust foundation. The fact that the echocardiograms were interpreted by board-certified cardiologists blinded to the ECG findings adds a lot of rigor to that ground truth.

Sam: The prevalence in that dataset is also something buyers will need to consider. A 34.8% prevalence of structural heart disease is quite high, which likely reflects the fact that these patients had a clinical reason to get an echocardiogram within 30 days.

Maya: Exactly. The real-world positive predictive value will heavily depend on where this software is deployed. If it is used on a broad population where the prevalence of SHD is much lower than 34.8%, the positive predictive value will drop accordingly, as their own estimates show.

Sam: So health systems buying this need to be prepared for the fact that a positive flag means the patient might benefit from further evaluation, but it is not a guarantee they have the disease. It is a notification software, just as the regulation name says.

Maya: Yes. For clinicians, the takeaway is that EchoNext adds a new layer of analysis to the standard 12-lead ECG. It might catch a signal of moderate aortic stenosis or left ventricular systolic dysfunction that would otherwise go unnoticed until symptoms became severe.

Sam: But it comes with strict boundaries. No pacemaker patients, and no using it to rule out disease.

Maya: To recap, the FDA cleared Pathwai's EchoNext, a machine learning software that analyzes 12-lead ECGs to flag suspected structural heart disease and six specific sub-labels. It demonstrated a sensitivity of 73.7% and a specificity of 75.2% for detecting SHD in a dataset of 102,219 patients.

Sam: It was found substantially equivalent to the Viz HCM module, acting as a notification tool to identify patients who may need further diagnostic evaluation, like an echocardiogram.

Maya: As always, the full FDA 510(k) summary is linked in our show notes. Read it for yourself.

Sam: And remember, this podcast is for informational purposes only and is not medical advice. Thanks for listening.