Can deep neural networks save your neural network? Artificial intelligence, sensors, and strokes

Each year, 15 million people suffer strokes, and at least a fifth of those are due to atrial fibrillation, the most common heart arrhythmia. Brandon Ballinger reports on a collaboration between UCSF cardiologists and ex-Google data scientists that detects atrial fibrillation with deep learning.

Each year, 15 million people suffer strokes, and at least a fifth of those are due to atrial fibrillation—the most common heart arrhythmia—where a blood clot forms within the heart due an irregular heart rhythm. With the release of Apple Watch and Android Wear, millions of people are now measuring their heart rate hundreds of times per day. Deep learning has created breakthrough results in speech recognition, computer vision, and machine translation—can it do the same for healthcare and potentially save lives? Brandon Ballinger reports on a collaboration between UCSF cardiologists and ex-Google data scientists that uses deep learning and Apple Watch data to detect atrial fibrillation. Beginning with a brief overview of cardiology, Brandon explains what both normal rhythm and atrial fibrillation look like when measured on their proprietary Apple Watch app, Cardiogram. Brandon then dives into how the team is applying deep learning to cardiology. He includes a very quick review of convolutional neural networks, dropout, rectified linear units, and all the other techniques that are now considered the “standard recipe” of deep learning; he also describes how the team uses a particular type of neural network—called a sparse autoencoder—to detect anomalous patterns of heart-rate variability. Brandon walks attendees through specific results, including sensitivity-specificity curves, and experiments the team has run to improve accuracy. Brandon concludes with some broader thoughts on the intersection of artificial intelligence and medicine, including tips learned by bridging the cultures of machine-learning research and clinical research, as well as some thoughts on how sensor data, the quantified self, and artificial intelligence will drive the future of healthcare—not just to calm the “worried well,” but with potentially life-saving algorithms.