The DeepHeart neural network can be used to detect early signs of diabetes with nothing more than just about any ordinary smartwatch, according to a study performed jointly by researchers at the University of California at San Francisco and the people behind health app Cardiogram. An interview with Cardiogram co-founders Johnson Hsieh and Brandon Ballinger revealed some details about how DeepHeart works and how it was created, as well as news that Cardiogram will be working on updates throughout 2018 that will integrate features based on DeepHeart into the Cardiogram app for users to take advantage of.
It’s worth noting that DeepHeart can be used to not only detect diabetes, but also hypertension, sleep apnea, and atrial fibrillation. That said, it was trained using real data from real people. 14,011 unique users of the Cardiogram app were recruited into a study, yielding 33,628 weeks’ worth of data that contained samples from both users in good health and users with the conditions noted above. The small variations in the way patients’ hearts worked could be spotted over time, and were analyzed and extrapolated into training data from DeepHeart. The training process was semi-supervised, with researchers essentially “telling” DeepHeart what some of the samples meant, and then having it figure things out from there. This wound up creating a neural network with 564,227 different weight points, or factors upon which it bases its analysis. After that, another study was conducted to ensure that the neural network was accurate in its findings when screening real people for symptoms of those conditions using data from the initial sample group. This study involved 12,790 weeks’ worth of data from an entirely new sample group. The final testing yielded roughly 85-percent accuracy in detecting the conditions in DeepHeart’s roster.
Once DeepHeart is integrated fully into Cardiogram, it will be able to collect data from the full measure of the app’s wide user base across both iOS and Android. As with anything else based on neural networks, this means that its performance and accuracy will vastly increase, likely in a fairly short period. The implications thereof are actually pretty wide; while it’s obvious that DeepHeart can improve the functionality of Cardiogram, it could also be used in professional healthcare applications like patient monitoring, or it could be modified to use different hardware and detect a wider range of conditions in the future.