As somebody who has both A. had a heart attack and B. is working on a machine learning project now, this is a subject that's very, erm, near to my heart ( no pun intended really).

Do you have an email address you can be reached at brandonb? I'd love to chat with you sometime if you're willing. I'm definitely interested in ways to apply machine learning to health-care.

Sure! Hit me up at [email protected].

This is really cool! It sounds like you're trying to determine whether you can identify AF out of data that's complicated by poor recording quality and regular activity? Is your longer term goal/business model to use this as a diagnostic tool (i.e. replace a Holter monitor)?

And since you offered to answer questions, I have a somewhat niche one. I'm an MD/PhD student doing my research in computational electrophysiological modeling, and I'm getting interested in the idea of using machine learning in medicine, specifically in signals analysis (i.e. EKG), and in time-series analysis (like heart rate variability). I'm teaching myself machine learning via a fantastic UW Coursera series that covers the basics (regression, classification, clustering, etc) with good mathematical rigor, but I'm not really sure where to go from there. What techniques would you suggest I look at? Would learning something like TensorFlow (I know Google has put out a course on Udacity) be a good idea? Obviously at some point, I just need to find some small projects to jump in on, but I want to build a decent skillset first if I can. Any thoughts would be appreciated!

If we succeed in training an accurate enough machine learning algorithm, I think the primary purpose will be to screen for asymptomatic atrial fibrillation, and then use a Holter or Zio Patch for a final diagnosis.

The main benefit is that you can prevent cryptogenic strokes, which are often caused by undiagnosed atrial fibrillation: http://www.nejm.org/doi/full/10.1056/NEJMoa1311376

For machine learning, I think you're on exactly the right path. The Udacity TensorFlow course taught by the Google Brain team is good; if you can complete that, you'll be as good as anyone at applying neural networks. The Stanford courses (http://cs231n.stanford.edu/, http://cs224d.stanford.edu/) are likewise great, as is Chris Olah's blog for intuition (http://colah.github.io/).

And down the road, if you're ever looking for an internship, definitely let us know! [email protected].

Having worked with some physicians on AF diagnostic instrumentation, I took away that an outsized difficulty in this diagnosis is patient compliance. One doc called it a "one shower" problem; very many patients would wear holters as prescribed until they needed to take it off, and then fail to don it again. Clearly this is an issue when hunting for an infrequent arrhythmia, and it seems that the comfort and ease of use of the device are inherent to the problem.

From what you're gathering, do you imagine an eventuality in which a medical device can take a form factor similar to these popular consumer choices and provide a signal of diagnostic quality? Or do you think PPG-style sensing is limited to a fundamental degree of uncertainty with respect to arrhythmias?

Yeah! Absolutely. In healthcare we call it "patient compliance," but it's really just usability.

My guess is that by the time we get to the Apple Watch or Android Wear 3 or 4, it'll include a built-in ECG sensor. There are some devices out there already (like Nymi) that have built ECG into a wristband -- if the user touches a wristband worn on one hand with a finger on the other hand, that completes a circuit through the heart and gives you a 1-lead ECG.

I think the downstream implication is that, just like smart phones replaced dedicated GPSes and music players, many medical use cases will become "apps" that run on general purpose wristbands. The big advantage is that since they'll be always-on and non-intrusive, we'll be able to catch conditions much earlier.

And if you're curious as to what atrial fibrillation looks like when measured on an Apple Watch in workout mode, here's an example (bottom-right corner): https://twitter.com/bballinger/status/714500654308306946

I downloaded the app, logged in with Facebook, and then went to mRhythmStudy.org only to find out that signing up for the study wants an email/password login. I don't have that since I used Facebook. And if I Sign Up for a new account, I don't see any way in the app to log out and back in with the new email/password account.

Oh sorry—that's confusing. You'll actually have two accounts. One for the Cardiogram app using your choice of Facebook / Twitter / Email, and one for Health eHeart study (our partners at UCSF) using email.

When you sign up for mRhythm on health-eheartstudy.org, create a new account, and then you'll be guided through a few survey questions and have the ability to link your Cardiogram account with your Health eHeart account. That link is what lets us do the research.

Does that make more sense?

Ok yeah that does make sense.

Wouldn't it make more sense to use a device that's not locked in to a specific vendor?

We do plan to support other devices like Fitbit and Android Wear. The reason we started with Apple Watch is that 1) the heart rate sensor is quite good, and 2) we can target millions of people with one platform (iOS / Watch OS).

I've been participating. I noticed that when I'm doing a workout my Watch"s heart rate monitor might drop to 65 bpm from 120, for example, then recover on another reading. Does this mean that the Watch isn't always accurate?

Thank you for participating! And that's exactly what we're measuring—how accurate is the Watch, and in what circumstances does it make errors? The anecdotal evidence is very conflicting, which is why we're comparing the AW to a few different clinical gold standards to get hard metrics.

Do you have any plans for integrating other sensors, such as many HRV capable bluetooth smart HR straps on the market?

edit: and other platforms, such as android

Yes! We plan to support other devices like Fitbit and Android Wear. My co-founder and I both worked on Android at Google, so we'd love to get back to Android Studio.

Looks awesome. Did you find HIPAA or the need for an extremely low error rate to be burdensome in building the app?

We're lucky in that information covered by HIPAA (e.g., the medical record) is handled by our academic partner at UCSF, the Health eHeart study (health-eheartstudy.org).

I think it's a bit of a misconception that you always need an extremely low error rate in medicine. Many of the existing tools are quite poor. CHADS-VASC, for example, is the primary score that cardiologists use to prescribe blood thinners, and its c-statistic is an... unimpressive 0.673. (0.5 is random, and 1.0 is perfect.) Given that the side effects of unnecessary blood thinners include brain hemorrhage, even a modest improvement to CHADS-VASC could save hundreds of thousands of lives.

So my message to data scientists considering going into healthcare is... please come! You don't have to achieve perfection to save lives. Even a mediocre statistical model would be an improvement in so many areas.

Could you point to a UCSF .edu page validating the mRhythm study and this app's connection to it?

Our academic partner at UCSF Cardiology hosts their study on their own domain (https://www.health-eheartstudy.org/). If you sign up through the links to Health eHeart provided at mRhythmStudy.org, you'll see a customized flow for the mRhythmStudy which lets you link your Cardiogram account.

Our principle investigator, Dr. Greg Marcus, also spoke briefly with with Fortune when the mRhythm study launched:

  https://fortune.com/2016/03/16/apple-watch-cardiogram-stroke-study/

  http://profiles.ucsf.edu/gregory.marcus