Ex-Googlers Build a Neural Network to Protect Your Heart

AliveCor's little EKG device has been around for a few years, but now it has a much smarter brain behind it.
KardiaMobileangledKardiaTA.jpg
Alivecor

The world knows no deadlier assassin than heart disease. It accounts for one in four fatalities in the US. Early detection remains the key to saving lives, but catching problems at the right time too often relies upon dumb luck. The most effective way of identifying problems involves an EKG machine, a bulky device with electrodes and wires. Even many portable machines like battery-powered Holter monitors, are unwieldy.

And so most people visit a doctor for an electrocardiogram. That, too, is no guarantee, because the best detection means being tested when a potential problem reveals itself. Otherwise, early signs of heart disease might go undetected. At-risk patients eager to keep an eye on their ticker might find a compact, easy to use, and EKG machine a good option. Like so many other gadgets, portable EKG machines are getting ever smaller---just look at products like Zio, HeartCheck, and QuardioCore.

The Kardia from AliveCor is about the width of two sticks of gum. Stick the $100 device on the back of your phone or slip it into your wallet, place a few fingers on it for 30 seconds, and you've got a medical-grade EKG reading on your phone.

That's impressive, especially for something so tiny. But the bigger story is what happens with the heart data it collects. AliveCor takes a big-data approach to safeguarding that essential organ inside your ribcage. The company uses neural networks and algorithms to identify signs of heart disease, an approach it hopes might change how cardiologists diagnose patients.

Beat Poetry

Changing cardiology with a 3-by-1.5-inch black wafer sounds dubious, but AliveCor convinced investors and the FDA. The Mayo Clinic and healthcare hardware manufacturer Omron recently announced a $30 million round of funding. "They are clearly a leader in this space," says Stanford geneticist Michael Snyder. The device's ease of use could lead to more frequent testing, and increased likelihood of early detection of heart disease. He says the investments from Mayo Clinic and Omron suggest "they realize that it is the future."

What's new here isn't the device, which AliveCor introduced a few years ago, but the neural network. The 40 people working for the company include former Google engineers who worked on Gmail, Google Maps, YouTube, and Google Photos. “We built a four-layer-deep neural network,” says AliveCor CEO Vic Gundotra, a former Google exec who came out of retirement to lead the company. “Four convolutional layers, two that are connected, 300,0001 parameters, and it takes about seven minutes for us to train."

Use the product for about a month, he says, and the technology builds your heart profile---a data-driven model that can detect anything amiss with your ticker. The neural network isn’t just looking at variances in heart rate, but in how the electrical system within your heart is firing. It’s smart enough to know if two people are sharing the same device, clouding the data pool in the process.

A.I. For the Heart

So far, AliveCor's tech can't be used to directly diagnose heart conditions. However, since it displays electrocardiogram data on a patient's smartphone, it can show if the person has an irregular heartbeat.2 One of the things an irregular heartbeat would indicate is an atrial fibrillation, a common early warning of stroke. That specialization facilitates the Kardia's small size. “The most common arrhythmia, the one that we need to detect, doesn’t need all 12 EKG leads,” Gundotra says. “And it’s not invasive at all, it’s just like brushing your teeth every day.”

Kardia-Pro-patient-file_screen-shot-1.jpgAlivecor

Still, the utility goes beyond that one test. Even if the system can't pinpoint a specific ailment, simply knowing something is out of whack can alert your doctor to take a closer look. Think of it as a "check engine" light for your heart. Gundotra offers an example from the device's clinical trials, when a longtime user of the Kardia received a warning.

“Thousands of EKGs, the neural network says 100 percent, this is the same person, and then boom,” Gundotra says. “In November, it says, ‘This does not look like the same heart anymore.’ This person developed right bundle branch block. Your heart has a bunch of electrical branches, and his right bundle got blocked. When would that be caught today? At your next visit if your cardiologist happened to do an EKG. We caught it the night his heart changed.”

That granularity may give patients peace of mind, but it can give doctors fits; imagine being inundated at all hours with EKG readings from patients. With that in mind, the new Kardia Pro software, which gives doctors a patient-by-patient dashboard of EKG data, alerts them only when a device records an abnormality.

So far, AliveCor’s back-end smarts only work with the company’s hardware, but the team is open to licensing its technology to others. It can see a day when you might find the technology in things like smartwatch bands, wearables, and even sensor-laden clothing. When that happens, checking your heart could be as easy as checking your temperature.

1Corrections: This story has been updated to correct the number of parameters in the neural network. Also, it was updated to clarify the role Kardia plays in disease diagnosis.