Posted October 20, 2016

In healthcare, prevention often falls into the “easier said than done” category: eat better, get more exercise, get more sleep, etc. On the other hand, due to a tremendous rise in new health-related data sources, there is an opportunity for a whole new approach to preventive medicine. Through its ability to interpret new data sources, machine learning is ushering in a new era of prevention. In some ways, then, this new era will all come down to the data. Perhaps two of the greatest surprises are that the data source may already be on your wrist — and with that data source, an app could save your life.


Consider the Apple Watch, Fitbit, Jawbone, or other wearables. Even today, they can yield a novel information stream that is relevant for your health. Millions already have these devices, so in many cases, the data is already there, just waiting for the right software to yield insights that are actionable, such as advising that issues that may seem minor enough to ignore to the wearer are in fact really quite serious. This suggests a future where wearables open the door for consumers to better take charge of their health.

These devices allow for constant monitoring, which is fundamentally different than a once-a-year visit to the doctor. Our lives are not determined by single data points. Consider your heart health today: you likely had a ton of variation from minute to minute, depending on whether you were catching the train, working out, or getting excited about some news. Sitting in the cardiologist’s office, one gets one clinical measurement (albeit a very proper one) of an anecdotal nature capturing just that one moment in time. On top of that, doctor visits are well known to be plagued by bias, depending on whether one gets relaxed or more anxious by that environment.

These devices do, however, currently have important limits: typically, any given data point is (at least today) not at a clinical level. But this begs the question, given that the wearables capture moments that would otherwise be missed, how can a less accurate device be useful? Is there some way that a less accurate device could be improved by having lots of additional measurements, which are themselves no more accurate? This is where modern machine learning comes in. New methods such as semi-supervised machine learning allows large but somewhat noisy datasets, combined with much smaller but much more accurate datasets (i.e., a few of the people also get full clinical tests). Together these allow for the simpler measurements to yield predictions comparable to the full tests. As long as one has some patients with those more accurate tests — such as in studies run by research teams — the knowledge gained from wearables can then be connected to data from more accurate measurements.


Here’s a simple example of semi-supervised machine learning in action. Let’s say we have some data run with the top clinical tests on patients. This gives us “labeled data” we have in addition to the wearable measurements, where the gold standard instruments tell us whether someone is healthy or not (circles vs triangles in Fig 2a). With just a few data points, however, it’s hard to tell where to draw the boundary between these two groups. By adding in additional unlabeled data (e.g., data from wearables, even without the clinical tests), one can better see the natural clustering of the data, allowing for a considerably more accurate and predictive boundary to be drawn. This helps us better know if someone is unhealthy. It also means that as we get more wearable measurements, the results get more predictive.

For Cardiogram, this isn’t just a goal, it’s a shipping product. Cardiogram’s app gets data from the Apple Watch today, and then uses datasets enriched by gold standard full measurements, thanks in part to their collaboration with UC San Francisco’s world class Cardiology experts. They’ve already collected more than 10 billion sensor measurements from more than 100,000 people with Apple Watches, allowing them to achieve a c-statistic (an accuracy measure) of above 90%. There’s already early evidence that this working. One recent review told the story in a nutshell better than any marketing expert could: when using the app, a recent user found there was an issue, called an ambulance, and was treated immediately, crediting the app with saving his or her life. As Cardiogram gets more users, their ability to make more accurate and more useful predictions will only grow.

The Cardiogram team reflects a new breed of entrepreneurs — Brandon Ballinger and Johnson Hsieh — who have expertise in both the computer science and biology domains, combining machine learning expertise from Google and Stanford with healthcare experience and the collaboration with UCSF. With so many unknown unknowns in the new world of combining computing with biology, teams in which the key members can span both spaces individually have enormous, critical advantages even over teams comprised of individuals with strengths in each separately.

I am very excited to announce that we’ve made an investment in Cardiogram to help them build this new product and hopefully help save many more lives.