Company Spotlight: Apple HTNF - Screening Blood Pressure From The Wrist

Sam: Imagine a consumer smartwatch passively checking your blood pressure in the background for a month, ultimately deciding to flag you for hypertension with a specificity of 92.3 percent. The implications for primary care waitlists are enormous.

Maya: Today we are unpacking exactly how that works by looking at the FDA 510(k) summary for the Hypertension Notification Feature, or HTNF, submitted by Apple and published by the U.S. Food and Drug Administration.

Sam: As always, you are listening to AI-generated voices, and we are summarizing a publicly available FDA clearance document, not reading from a company press release.

Maya: Right, this is a Company Spotlight episode, so we are keeping it strictly to the facts cleared by the regulator. And this document is fascinating for both clinicians and tech developers because it outlines exactly how a ubiquitous consumer device steps over the line into medical screening.

Sam: Let's start with the basics. What exactly is this feature, and who is it actually for?

Maya: The HTNF is a software-only mobile medical application. It analyzes photoplethysmography data, which is the optical PPG sensor on a compatible Apple Watch, to identify patterns suggestive of hypertension. The feature then surfaces a notification to the user.

Sam: And the indications for use? Because there are always very specific guardrails in these FDA summaries.

Maya: Exactly. It is intended for over-the-counter use by adults age 22 and over who have not been previously diagnosed with hypertension. It is not intended for use during pregnancy.

Sam: So it's purely for undiagnosed adults. What if someone is already on blood pressure medication and wants to see if it's working?

Maya: The document is extremely clear on that. It states it is not intended to replace traditional methods of diagnosis, to monitor hypertension treatment effect, or to be used as a method of blood pressure surveillance.

Sam: Okay, so this is a passive, opportunistic screening tool to nudge people into the healthcare system, not a management tool for people already in it. How does the software actually figure out someone might have high blood pressure just from a light sensor on their wrist?

Maya: The summary breaks down a machine-learning algorithm with two main components. First, there is a deep-learning model. This was developed using self-supervised learning on a large-scale unlabeled dataset of Apple Watch sensor data from over 86,000 participants.

Sam: Wait, really? 86,000 participants just for the base model? That's a massive scale.

Maya: Yes, but remember that was unlabeled data to extract generalizable characteristics of the PPG signals. On top of that, they trained a linear model to actually classify the risk as hypertensive or non-hypertensive. For that part, the development dataset included watch sensor data and home blood pressure reference measurements from over 9,800 participants.

Sam: So the watch scores the PPG signals, and then what happens? Does it alert you immediately if you have a spike?

Maya: No, it's not instantaneous. The software on the watch gives each qualified signal a score. But the software on the iPhone is what aggregates those risk scores and looks for patterns suggestive of hypertension. If it finds them, it surfaces the notification. It operates over multiple days, and the document notes that it will not surface a notification if insufficient data is collected.

Sam: This brings us to a really interesting part of the FDA process. This is a 510(k) clearance, which means Apple didn't have to prove this was a totally novel, first-of-its-kind medical device. They just had to prove it was substantially equivalent to an existing, legally marketed device, which we call a predicate.

Maya: That's right. And for anyone building AI in healthcare, the choice of predicate here is a brilliant lesson in regulatory strategy. Apple used a device called Viz HCM, manufactured by Viz.ai, Inc.

Sam: Let me stop you right there. HCM stands for hypertrophic cardiomyopathy. That's a structural heart muscle disease. And Viz HCM uses 12-lead ECGs, right? How is a watch looking at optical blood flow for hypertension substantially equivalent to a 12-lead ECG looking for a thickened heart muscle?

Maya: It sounds like a stretch, but under FDA regulations, they both fall under the same product code and regulation name: Cardiovascular Machine Learning-Based Notification Software. They are both software-only mobile medical applications intended to identify a single condition based on non-invasive physiological inputs and suggest the likelihood of a cardiovascular disease for further follow-up.

Sam: So the FDA isn't looking at whether the disease is the same, or even if the sensor is the same. They are looking at the mechanism of action—software notifying users of a potential cardiovascular issue based on a machine learning analysis of physiological data.

Maya: Exactly. The FDA summary clearly lists the differences: Apple's HTNF is over-the-counter for lay users, passive, and looks at PPG over multiple days for hypertension. Viz HCM is a prescription device for clinicians, on-demand, and looks at 12-lead ECGs for hypertrophic cardiomyopathy.

Sam: And the FDA concluded those differences do not raise new questions of safety or effectiveness. That opens a lot of doors for digital health companies looking for predicates. But a clearance is only as good as the clinical data backing it up. Let's talk about the pivotal clinical validation study. What are the real numbers?

Maya: They enrolled a total of 2,229 subjects without a previous diagnosis of hypertension. Participants wore the watch for 30 days and took reference measurements using an FDA-cleared home blood pressure monitor. Out of the 2,229 enrolled, 1,863 provided at least 15 days of usable data for the primary endpoint analysis.

Sam: Okay, 1,863 usable subjects. Hypertension in this study was defined by the American Heart Association standard, which is an average systolic blood pressure of 130 or greater, or diastolic of 80 or greater. How well did the algorithm actually perform?

Maya: The overall sensitivity was 41.2 percent, and the specificity was 92.3 percent. They also reported a positive predictive value, or PPV, of 70.9 percent based on a prevalence of 31.4 percent in the study.

Sam: Let's translate those numbers for the boardroom and the clinic. A sensitivity of 41.2 percent means the algorithm is missing more than half of the people who actually have hypertension. That feels low for a medical device.

Maya: It is low if you treat it as a diagnostic test, which is why the indications strictly state it is not intended to replace traditional methods of diagnosis. The summary explicitly warns: the absence of a notification does not indicate the absence of hypertension.

Sam: So from a liability and safety perspective, they are saying, 'do not assume you are healthy just because your watch is quiet.' But let's look at the flip side. A specificity of 92.3 percent is quite high. That means if the watch does ping you, it's very rarely a false alarm.

Maya: Exactly. They tuned this algorithm to be conservative. They want to avoid flooding primary care offices with millions of panicked, healthy people. When they looked at stage 2 hypertension, which is a more severe elevation, the sensitivity went up to 53.7 percent. And the specificity for purely normotensive subjects was 95.3 percent.

Sam: That makes total sense for a consumer deployment. You don't want false positives when you're shipping to millions of wrists. Did they break down the performance by demographics? PPG optical sensors have historically struggled with darker skin tones and obesity, because you're literally trying to shine a light through tissue to see blood volume changes.

Maya: They did conduct pre-specified demographic subgroup analyses adjusted for age, sex, BMI, race, and blood pressure. For skin tone, comparing Fitzpatrick Skin Tone I-IV against V-VI, the risk ratio confidence intervals included 1.0. However, for race, the specificity risk ratio comparing Not White versus White was 1.04, with a 95% confidence interval of [1.01, 1.07].

Sam: Meaning there was a statistical difference for that specific race comparison? Did that raise any red flags?

Maya: Not according to the summary. It states that comparisons of sex, race, and skin tone suggested no clinically meaningful difference after covariate adjustment. For example, they noted the Asian subgroup was younger, with a mean age of 43.0 versus 50.9 years, and had a lower BMI of 27.7 versus 30.9 kg/m2. After covariate adjustment Asian performance characteristics were on par with non-Asian participants.

Sam: Did any subgroups show a drop in performance?

Maya: Yes. They observed higher sensitivity and lower specificity in subjects with an older age, defined as 60 years or older, and a higher BMI, defined as greater than 30. However, the FDA summary states this result was deemed clinically acceptable due to the higher risk nature of these subgroups.

Sam: That's an interesting regulatory tradeoff. The test is less specific for older people with a higher BMI, meaning more false positives in that group. But because those people are at higher risk for cardiovascular disease anyway, sending a few more of them to the doctor for a blood pressure check is considered an acceptable clinical risk.

Maya: Exactly. They also looked at longitudinal performance over a two-year period, breaking data into six non-overlapping 30-day windows. They found that the specificity for non-hypertensives remained high over the long term, at 86.4 percent for a subset of 187 subjects.

Sam: What about the people who got false notifications? Did their watches just keep pinging them every month for two years?

Maya: According to the summary, the majority of participants without hypertension who received a notification did so in the first month, and the number of new notifications within this group tended to decrease month-over-month. And for the 121 non-hypertensives who had strictly normal blood pressure, the specificity over two years was 92.5 percent.

Sam: Let's pivot to the backend, because there is something in this document that every software developer in medtech needs to pay attention to. The PCCP.

Maya: Yes, the Predetermined Change Control Plan. This complies with a relatively new framework under the Food and Drug Omnibus Reform Act of 2022.

Sam: Right. Historically, if you changed a cleared medical software algorithm, you often had to submit a whole new 510(k) to the FDA. The PCCP lets companies outline exactly how they plan to tweak the algorithm in the future, and if the FDA agrees to the plan, the company can make those updates without a new submission.

Maya: That is what Apple did here. The document includes a table of proposed modifications under the PCCP. It includes revising the PPG signal input characteristics, incorporating demographic data input types for the ML model, implementing quantization, retraining the model, and modifying the logic for surfacing a notification.

Sam: That is basically a license to continuously improve the product. But there's a catch, isn't there? AI models can't just learn on the fly.

Maya: Correct. The summary explicitly states: The PCCP does not include provisions for implementation of adaptive algorithms that will continuously learn in the field. All algorithm modifications will be trained, tuned, and locked prior to release of the software.

Sam: Locked models. So they gather real-world data, take it back to Cupertino, retrain the model, lock it, and then push an update to your watch. And how do they validate those updates before they push them?

Maya: The PCCP defines two test methods. There is a Single Watch Test Method to establish non-inferiority of the updated performance compared to the initial release. And there is a Dual Watch Test Method to evaluate the updated software on an updated hardware platform.

Sam: Wait, dual watch? Do they literally make study subjects wear two Apple Watches?

Maya: They do. The data collection approach requires subjects to simultaneously wear one watch on each wrist and undergo daily data collection. The summary notes that they swap the watches on day 15 of the 30-day study period to minimize physiological or behavioral biases from wrist wear.

Sam: That is incredibly rigorous, and also a very funny image to imagine someone walking around with a watch on each wrist for a month to validate a software update. But it highlights the core challenge here: Apple is using a general purpose computing platform as a medical device.

Maya: Exactly. The FDA summary even notes that because it's a multiple function device product, they had to assess the impact of the general purpose platform on the HTNF software. But they deemed it acceptable, just as they did for Apple's previous clearances for irregular rhythm, atrial fibrillation history, and sleep apnea.

Sam: So, wrapping this up. Apple has successfully cleared a hypertension notification feature for adults 22 and over. It's a passive, multi-day monitor with a specificity of 92.3 percent, meaning few false alarms, but a sensitivity of 41.2 percent, meaning it will miss cases and shouldn't be used to definitively rule out high blood pressure.

Maya: And from a business perspective, they achieved this through substantial equivalence to a different cardiac notification software, and secured a Predetermined Change Control Plan to allow future locked-model updates without repeated FDA submissions.

Sam: It's a massive shift in how we think about screening. Millions of people will soon have a passive monitor on their wrist, and health systems need to be ready for the inevitable influx of asymptomatic patients walking into clinics saying, 'My watch told me to come here.'

Maya: As a reminder, you can find a link to the full FDA 510(k) summary for the Hypertension Notification Feature in our show notes. And a final disclaimer: this podcast is for informational purposes only and is not medical advice.

Sam: Thanks for joining us on AI in Medicine - Smart Summaries. See you next time.