Deep Learning in Cardiology

Thirteen years ago in my first anatomy class of Med School, the instructor asked us to make sure our learning is “deep.”

“You need to memorize the names of every single tiny nerve and muscle, because they all will be on your exam. One day you could be surgeons and if you cut out the wrong structure, you can kill someone!” he asserted as we all stood there in fear.

Later in Med School, we were told that half of what we’re learning will be wrong by the time we practice. But one thing we were not told is that the way we learn and the entire premise of what makes a good doctor would also change. For example, “deep learning” itself now means something different to me and to most healthcare professionals. If you are reading this article so far, then you are likely to have seen the term floating around in medical journals.

Deep learning is a type of machine learning in which the computer uses multiple layers of processing to extract features from otherwise vague data input, such as an ECG or a slice of an MRI.  Each layer uses outputs from the previous layer. Through deep learning, the computer simulates the neural network of the brain and is able to learn and make sense of abstraction.

Working at the Broad Institute of Harvard and MIT allowed me to recently be part of a team that uses deep learning to solve important problems in cardiology.  Over several weeks, cardiologists, scientists, and machine learning experts worked in teams to train computers on deep learning models so that they understand data such as medications, ECGs, genomic architecture, and cardiac MRIs of tens to hundreds of thousands of people.

The insights I gained were incredible.  Just like medical students perform better on their exam if they learn “deeper,” the longer you train a computer model, the more it learns and the better it performs in predicting – but it does so at much faster rates than any doctor could ever match. For example, in only two days, we trained a computer to read the ejection fraction from a cardiac MRI as good as a doctor would. Using the MRI, the computer could also predict with reasonable precision the presence of hypertension and coronary artery disease, without knowing anything else about the patient. The power of computer vision is beyond imagination. While it could take you a full day to read 100 ECGs, a well-trained deep learning model could read them in only few seconds. It could also identify patterns in the data that the human eye could not discern, which might or might not be biologically or clinically relevant.

As data availability and computing power continue to grow, we will be seeing more and more applications of deep learning in cardiology.  While we do so, we should stay mindful that human supervision and our role as doctors in charge of our patient’s health is more important than ever. This requires us to understand how computers work and how those models are built through working with multidisciplinary teams. If we do this right, we can probably do less deep learning ourselves by delegating to computers, and gain a whole lot of extra time that we can invest in taking care of our patients.