How Apple built hypertension notifications for Apple Watch

February is Heart Month, so it’s appropriate to speak with the team that built the recently introduced hypertension notifications system for watchOS 26 and Apple Watch. 

I spoke with Apple’s Steve Waydo, director for health sensing, and Dr. Rajiv Kumar, physician-researcher, who offered a glimpse into the science and decisions behind their lengthy project to give smartwatch users an actionable and reliable tool to track this aspect of heart health.

Waydo led the long development of the hypertension notification feature. “The idea goes way back to not all that long after we launched the first Apple Watch,” he said. “We had this device collecting physiological data on users all the time. This hadn’t existed before. We saw it as a new and unique opportunity.”

But first, Apple needed to develop new sensor capabilities, assemble world-class technical and clinical expertise, and build accurate and effective machine learning tools. Apple wanted its solutions to be grounded in science, so it also launched a large-scale heart health study with the University of Michigan. 

Understanding Hypertension

Hypertension is a state of chronic high blood pressure. Each time your heart beats, it moves blood out of the heart and into your blood vessels. When blood pressure is high, there’s a lot of back pressure so the heart must beat harder than usual to get the blood out. That, over time, is called hypertension. The problem with this condition is that it’s totally asymptomatic, which is why it is seen as a silent killer. 

More than 1 billion people have this condition, and nearly half of all US adults suffer from it. Yet, around half of them don’t know they have it. 

That’s why Apple’s tool is important; it could help people identify the condition and take steps to manage it. “So much of our health is invisible even to ourselves, and one of the biggest barriers to better health is just simply not knowing what’s going on,” said Waydo.

Apple’s plan is to harness the power of wearable data to help surface conditions such as this one, which otherwise might not be easily managed by any of us.

Machine learning, data, and context

The data an Apple Watch provides differs from most test data because the device is worn all day, almost daily. That means the information it gathers changes over time, which helps identify deep health insights. What makes the information more actionable is artificial intelligence, which helps the device itself surface useful insights based on the data it can track. 

Kumar explained how Apple developed a machine learning system to combine that personal data with real world information drawn from the Apple Heart study. The latter helped Apple understand, “what the signals look like, what they look like across a person’s life and in a variety of circumstances and break the raw sensor data down into thousands of independent factors that we can quantify.”

Apple also leaned into supervised learning data, in this case information derived from both sensor data and ground truth. This is the kind of information generated by Apple’s work with the University of Michigan. The beauty of the combination is that Apple can see how sensor data correlates with scientific data. Machine learning models can then analyze the personal data and contrast it with sensor data across thousands of factors to identify a person’s hypertensive status.

You can learn more about how Apple’s system works by reading the company’s extensive white paper on the topic.

Apple Watch, a wearable doctor

“These machine learning tools are a key enabling technology, because with something like hypertension, the way it manifests in our signals is extremely subtle,” said Waydo. “It’s really subtle features of the actual shape of the signal that we get off the sensors…. We’re looking at much more subtle signals that correlate with high blood pressure, because those signals tell us something about how your blood vessels respond every time your heart beats. So, we apply these machine learning techniques to millions of data segments.”

“I’ve been at Apple for 13 years, so I’ve been here along this whole journey,” says Waydo. “And these same kinds of tools make it possible for your watch to track your activity, understand if you’re walking, or swimming in a pool, estimate how long you spend in any sleep stage, identify when you take a fall, so that it can connect you with emergency services. So, we’re using machine learning tools all over the place.”

In each case, Apple finds that it is important to look at how a person’s data evolves over a long period of time, as opposed to just giving a notification based on one moment.

The art of noise

The phrase “garbage in, garbage out” does a lot of work in the AI age, but Apple’s experts had interesting insights into the nature of data noise. “You know, we are processing vast amounts of data to develop these features,” says Waydo. That means the algorithms must figure out how to grade the data they pick up. 

Apple, which supplements the data with research acquired from large-scale, real-world studies, found that building in support for “messiness” can make for better results. “Having that data set that actually that has messiness and realism to it is very important for coming up with signals that are more than a research curiosity and can really apply to, you know, actual people using our devices in the world,” Waydo said.

In AI, failure builds success

Getting the system to work was a long process of iteration and repetition. Apple’s teams built bigger and better data sets, revised their algorithms, and kept improving what they had built until it became ready to roll it out into the world.  “We rinse and repeat that process for weeks or months or years until we have something we’re happy with,” said Waydo.

Apple

The team also gets excited when things don’t work. “We may find use cases or particular kinds of users or particular scenarios where we get a lot of false positives or where we get no true positives. And that tells us where to go in order to improve the algorithm and iterate on the algorithm, and usually that means getting more data that captures those use cases that we can incorporate into our machine learning training.”

As part of the work, Apple’s teams also looked closely at demographics. The intention is to ensure that age, sex, or race don’t impact the performance of the systems Apple provides. Apple is a global company that ships products to a lot of people. It’s solutions have to work for everyone.

What it isn’t

The feature isn’t intended to be a complete replacement for regular check-ups. Recent reporting that not every case of hypertension will be picked up is correct, and reflects the balance the developers had to reach to create a system they could ship. That’s because the team realized that training the algorithm to be more sensitive would diagnose more cases, but at the cost of more false positives. 

The danger of false positives is that people stop listening. After all, if you are given health notifications by your device you must be able to trust its accuracy. No one wants to be given false information.

Waydo explained the conundrum: Should Apple aim for 100% sensitivity when it means the system will have a lot of false positives? Or should it aim to build a system that minimizes those? That’s why Apple had to achieve a balance.

“We weigh our work very heavily towards trying to manage false positives and trying to make sure that when we do notify someone of a potential issue, whether it’s hypertension or a regular rhythm or any of these other things, that notification is really trustworthy. And sometimes that means that there are cases that we can’t catch, because if we catch this additional set of cases, we’re also going to end up catching a bunch of people who don’t need the notification. And that undermines the utility of the whole thing,” Waydo said.

Apple recommends that any Apple Watch user receiving a hypertension notification check and log their blood pressure and visit their doctor.

Fan mail

Apple’s rock solid commitment to privacy means its teams can’t track how successful its systems are in the field because it never sees that kind of information coming from personal devices. But the team does get letters from users, medical practitioners, and others who have been affected

“I love hearing from clinicians who say they met someone who otherwise wouldn’t have known or who wouldn’t have come in, and it’s really changed their lives,” said Dr. Kumar. “Each of our features, whether it be in women’s health, hearing health, or heart health — they’re all based on science, must be actionable and absolutely built with privacy at the core.”

And each time someone gains better insight into their own health, they become better equipped to improve the health decisions they take in the future.

Please follow me on Twitter, or join me in the AppleHolic’s bar & grill and Apple Discussions groups on MeWe. Also, now on Mastodon.How Apple built hypertension notifications for Apple Watch – ComputerworldRead More