Machine Learning to Help Individuals with Severe Congenital Heart Disease
Machine learning is a field of computer science that uses statistical techniques to give computer systems the ability to "learn;” that is, they progressively improve performance on a specific task involving data. This article describes one of the first applications of machine learning to study congenital heart disease. It shows the great potential of these techniques in general biomedical research.
Approximately 1 in every 2,500 babies born in the US has a complex form of congenital heart disease called “tetralogy of Fallot” (TOF). Surgeons have become very skilled in “fixing” these hearts (more than 97% of children survive into adolescence and beyond). However, the overall health and life expectancy can vary widely as these patients become adults. While some individuals with TOF have become Olympic gold medalists, many others struggle with heart failure or other severe symptoms. Therefore, a major challenge for physicians is predicting which patients will experience worsening heart function in order to provide therapy.
To address this problem, our Department of Imaging Science & Innovation worked with world leading experts in pediatric cardiology from Boston Children’s Hospital to apply machine learning. Machine learning provides advanced methods for detecting patterns in data. In other words, by studying data from many patients, computers are able to “learn” the characteristics or patterns of similar patients and use this knowledge to make predictions. With the growth of computing power in recent years, a computer can now perform these pattern recognition tasks from large, complex data sets better than a human can. These strengths are why machine learning techniques are becoming widely used in many applications from healthcare to marketing and social networking.
In our study, we combined imaging, surgical and electrophysiologic data from roughly 150 patients with TOF to learn how to predict worsening heart function over time. For these patients, we had measurements from multiple time points (from at least 6 months up to 6 years), which we could use to help “train” the computer. Previous research using traditional techniques have generally failed to detect such patterns in TOF. However, applying machine learning to this problem produced more useful feedback. We found that, rather than one single marker, a complex pattern involving several markers from heart imaging and other clinical information, such as surgical details and patient age, emerged. For example, this pattern was able to separate patients (although, not perfectly) whose heart function got worse over time versus those whose heart function stayed the same (see picture).
Now that this pattern has been identified, physicians at Boston Children’s Hospital and Geisinger will be able to use this information for new patients they encounter to help inform their treatment. For example, patients who are predicted to be at risk for worsening heart function may benefit from additional medications, an implanted medical device to help control heart rhythm, or even additional surgery.
We are proud of the fact that leading experts in the field have positively reviewed this work. In particular, Dr. Cedric Manlhiot stated in his recent editorial that "Clinical studies, such as [this] are an excellent example of how to properly introduce machine learning algorithms in medicine that demonstrate clinical value in a realistic manner and we should hope that others will follow this example in the near future.”
Find more information about the study here.
About the Author
Brandon K Fornwalt, M.D., Ph.D., directs the Cardiac Imaging Technology Laboratory at Geisinger, which is a group of engineers, researchers and physician-scientists within Geisinger’s Department of Imaging Science & Innovation.
For more on this subject or Geisinger’s work in this area, please contact David Stellfox in Corporate Communications, Research, at firstname.lastname@example.org or 570-214-6549.