The Apple Watch learned to detect falls using data from real human mishaps
Apple explained how they collected the data to create the novel feature.
The Apple Watch in 2018 is more than just a time-telling wearable that taps your wrist when you have a message or counts your calories as you walk around: it’s evolving into a type of consumer medical device. Not only can it take an ECG to possibly detect a heart rhythm called atrial fibrillation (although that feature isn’t enabled yet), it can also detect if you’ve fallen down, and then call emergency services.
It’s a feature that has the potential to help the general population: according to the National Institute on Aging, more than one-third of people over 65 tumble annually. If the wearable detects that someone has wiped out, it can call 911 if the person doesn’t respond. For users over 65, Apple automatically enables this feature.
The Watch detects falls using a similar mechanism with which it tracks another complex form of motion—swimming. In the water, the timepiece uses its accelerometer (which measures changes in motion) and the gyroscope (which detects the rate of rotation along three different axes) to detect which stroke the wearer is doing. It must be able to recognize the difference between freestyle and butterfly, for example—which look similar, sensor-wise, although they burn calories at different rates—and to notice when the swimmer has changed directions at the end of a lap.
To teach the watch to monitor swimming, Apple collected data from hundreds of swimmers. And to create the algorithms for detecting falls, they got their data from real-world, unplanned, gravity-fueled interactions between humans and the ground.
To do that, they gathered data from people wearing Apple Watches (running customized software) in a movement disorder clinic, assisted living facilities, and friends and family of Apple employees. That study involved more than 2,500 people and ultimately included more than 250,000 days of data. The data came from real falls—like a spill off a ladder, a trip on a walk, or just a fall while getting dressed (thanks, pants).
This type of information is much more valuable than the kind of readings Apple would have gotten if they’d asked someone—a stuntman, say—to purposely fall; those actions may not be representative of a real-world spill. Their studies also included data on what normal motion looks like, to differentiate between a fall and actions that could resemble a fall. The company wanted to make sure that activities like swinging a tennis racquet or flopping down on a bed don’t register as life-threatening spills.
“We learned that with falls, there’s this repeatable motion pattern that happens,” Jeff Williams, Apple’s chief operating officer, said during the company’s keynote in September when announcing the feature. “For example, when you trip, your body will naturally pitch forward, and your arms will go out involuntarily to brace yourself. However, if you slip, there’s a natural upward motion of the arms.”
Of course, both swimming and a fall involve your entire body, but the watch is measuring only the three-dimensional trajectories of your wrist through space to infer what’s happening.
The key sensors that make this possible are the accelerometer and the gyro. The accelerometer in the Series 4 gathers eight times more data per second than the previous version, and it can measure a higher amount of G forces (32 Gs, up from 16 Gs). When someone swings their hands during a fall, that creates G forces (the “G” stands for “gravity”), but the greatest spike in Gs happens when their hand smacks the ground. Because the accelerometer can now capture up to 32 Gs of force—which is a lot—that means that Apple can register the big impact spike a hard fall can create, as opposed to the sensor’s measuring abilities maxing out at 16 Gs.
Then there’s the gyro, which is now more power-efficient—an important point, since Apple needs it to be powered on to monitor for falls throughout the course of the day. (It can switch off automatically if your wrist is relatively still on a table.) That gyroscope measures rate of rotation, and to visualize the different ways it does this, picture an axis going horizontally across the screen (the X axis); another one going vertically up the display (the Y), and finally a third sticking straight out through, and perpendicular to, the screen (the Z).
If you’re wearing a watch right now, hold it in front of you and tilt it towards you: that’s a rotation along the X axis. Now, put your palm straight down on a table: the Z axis is shooting straight up to the sky when the screen is parallel to the table’s surface.
Fall detection requires data from a combination of these two sensors, the accelerometer and the gyro. A tripping fall may result in an impact recorded by the accelerometer, and the screen itself may then have a distinct orientation—likely vertical— compared to the ground. That’s because your palms may end up flat on the ground, your wrist and forearm may be vertical and the watch screen with it, and that X axis is thus oriented vertically, too.
It’s important to note, too, that the SOS function baked into this feature won’t be able to call 911 if you don’t have the cellular version of the watch and you’re far away from your phone, because the non-cellular timepiece needs your handset to be in Bluetooth range to make a call through it. (Here’s more info from Apple about the process.) In other words, if you go for a run and want the alert-the-authorities portion of fall detection to work, bring your phone with you, too.