AI + Consumer Wearables + Atrial Fibrillation = Another Medical Breakthrough

March 30, 2018

Photo by Crew on Unsplash


While the non-stop discussions regarding the future of AI and humanity have not come to any conclusions, AI is steadily making its way into various spheres of life. One of such spheres is medicine, where it has already been labeled ‘the stethoscope of the 21st century’. For starters, let’s remember that the original stethoscope also took a while to be accepted by the medical professionals. Today even robot-assisted surgeries are no longer a rarity, in fact, in 2016 there were already 4000 of such robots throughout the world, taking part in an amazing number of operations – 750,000.

Today we’d like to focus on a very specific and very important aspect of the conjunction between medicine and artificial intelligence. It does not seem nearly as much of a breakthrough as robotic surgeries, but it can potentially save even more lives. We are referring to the ability of consumer wearables to detect atrial fibrillation, a dangerous and common condition. Atrial fibrillation is defined as an irregular and often fast-paced heartbeat, which puts a person at an increased risk of stroke and heart failure. Physically it involves the two atria (upper chambers) beating out of rhythm with the ventricles (lower chambers). It is the most common heart rhythm abnormality, and is often diagnosed when it has already caused major damage. Certain types of atrial fibrillation do not require treatment, but some are life-threatening and lead serious to complications. Diagnosing this condition has required an electrocardiogram to distinguish it from arrhythmia and examine the extent of the problem – until recently.

Cardiogram leads the way

In 2017 UC San Francisco and Cardiogram have published a series of research findings on the accuracy of medical condition diagnosis by consumer wearables. These conditions included hypertension, sleep apnea and diabetes. The new Cardiogram’s peer-reviewed clinical study presented in the respected JAMA Cardiology in March 2018 found Apple Watch and Android Wear to be highly accurate in detecting atrial fibrillation. The deep learning model used by the company had processed an incredible amount of data, estimated at more than 100,000,000 units. The uniqueness of the learning model presented is that it required far less data to be built. In this case, only 6,338 ECGs were necessary to construct the 8-layer model, and this is a true milestone for medical model-building, since previous deep learning models required far more data.

One of the issues about the study that need to be addressed is that it has drawn on the patients with a pre-determined risk of atrial fibrillation, while further research is required to determine whether this model can pro-actively detect the problem in those who have no previous treatment or diagnostic history.

Consumer wearables generate incredible amounts of data, which is definitively impossible to process for human doctors and scientists alone. This is where AI comes in, and Cardiogram is invested in continuing to improve DeepHeart, a deep learning neural network that’s already been proven highly efficient in rigorous clinical trials, and is bound to become even more effective as it continues learning from the data fed into it. It’s a semi-supervised system, and requires approximately 10 times less labeled data that traditional deep learning methods.

This is a significant step forward in clinical research, which should reduce health care costs and dramatically improve patient outcomes, and it’s only one of the first steps in this direction. We’re on the verge of a revolution in medicine, and we’re looking forward to reporting on each of the puzzle pieces that combine into a more and more efficient image of health care of the future with each passing day.

Leave a Comment

Your email address will not be published.