Seminar Series Synopsis: How Machine Learning May Help Us Study a Prevalent Heart Condition

Several years ago, Anthony Philippakis, cardiologist and chief data officer of the Broad Institute of MIT and Harvard, first learned of an artificial intelligence system that could predict patients’ age and sex based off of their electrocardiogram.

“As a cardiologist, when I first saw this result, I couldn't believe it,” Philippakis said. “No cardiologist I know can do that or even come close to it.” If an AI could identify those traits with such accuracy, what types of clinically useful information could it potentially find?

Philippakis shared the anecdote at a September 29 presentation hosted by Caroline Uhler at Jameel Clinic for Machine Learning in Health at MIT. Uhler is a Jameel Clinic Principal Investigator and MIT associate professor in Electrical Engineering & Computer Science. Her talk delved into the potential for machine learning to identify causal mechanisms of disease at the cellular, tissue, and organism levels.

Machine learning is a type of artificial intelligence in which a computer system improves its accuracy at a given task over time as it experiences more data. It’s not so different from how humans are better at flashcards after running through them several times.

Uhler’s project at Jameel Clinic is entitled “Representation Learning to Elucidate the Disease Mechanisms in Atrial Fibrillation.” Atrial fibrillation is a condition characterized by irregular heartbeat. It can lead to blood clots, or even stroke or heart failure. It is the most common form of treated heart arrhythmia, accounting for about 454,000 hospitalizations in the United States every year.

Cardiomyocytes are a type of muscle cell responsible for pumping the heart. For the heart to function properly, cardiomyocytes should contract in a regular, coordinated manner. How exactly do they do that? This is one of many questions about cardiomyocytes that scientists hope to address at the cell, tissue, and organism levels, Uhler explained in her talk.

“At every one of these levels, you have a data integration problem,” Uhler said “And you have actually very different data integration problems with each.”

Machine learning algorithms, adept at sifting through and finding connections within tremendous amounts of data, could help scientists resolve these data problems.

All of the cells in a person’s body, whether a brain cell, a skin cell, or a muscle cell, carry that person’s unique genetic code. Yet, a skin cell behaves very differently than a brain cell for two main reasons: how the genetic material in these cells is structured, and how it is expressed. The structure is the way that the DNA is packed and arranged within the cell. The expression is which genes are turned “on” in that type of cell. Both aspects are very important to the cell’s function, and they work together to determine the cell’s behavior.

“The challenge here is that you can never measure them together, because actually getting a [structural] image means killing the cell, and getting the expression means killing the cell,” Uhler said. “So you'll never be able to measure both in the same cell. So you need to have these unpaired datasets you need to somehow be able to translate and integrate between.”

This is one place where machine learning may be able to help scientists learn more about the cardiomyocytes behind atrial fibrillation. Think of expression and structure as two different languages, like French and Spanish. If you input a Spanish word into an online translator, it can spit back out the French term. Similarly, researchers may be able to devise a machine learning system capable of translating between specific cells’ structure and expression. Machine learning may be able to help scientists predict a cell’s structural problem based off of its expression, and vice versa. This provides researchers with a better sense of how a given cell functions from a more complete structural and expressive standpoint.

Scale a similar line of thinking up to the patient level. Electrocardiograms (ECGs) are a type of medical test that are less expensive and more accessible than MRIs. What if a machine was taught to correlate ECG results with MRI results, so that the information gleaned from an ECG was similar to that from an MRI?

“What people are starting to realize is that we can actually recover a lot of information and see things that generations of clinicians have missed,” Philippakis said. As an example, cardiac amyloidosis is a treatable heart condition which can be difficult to diagnose. “If your ECG spit out, ‘this patient likely has cardiac amyloids,’ that would be really useful [for physicians].”

The talk was part of Jameel Clinic’s weekly seminar series, taking place virtually and on MIT’s campus, Wednesday evenings from 4-5 p.m. EST. Jameel Clinic is the epicenter of AI and Healthcare at MIT, aiming to develop AI technologies that will improve the future of healthcare, including early diagnostics, drug discovery, care personalization and management.