Can machine learning predict bleeding during heart surgery?
Bringing together man and machine, a team of researchers is studying how machine learning could be used to better predict the probability of a patient bleeding during and after heart surgeries. Their findings have set the stage for providing doctors, researchers and clinicians a vital, data-driven tool to help them quickly determine whether a patient is at risk.
Analyzing previously collected data from the American College of Cardiology’s National Cardiovascular Data Registry, which includes more than 3 million procedures conducted across the United States, the researchers investigated how applicable and effective machine learning would be to identifying high- and low-risk patients.
“We need to be able to really explain what’s going on and understand what it means for somebody to be high risk or low risk (of bleeding) so that treatment decisions can be made off of that,” said Bobak Mortazav, assistant professor, Department of Computer Science & Engineering.
This could be the difference between life and death in the operating room. And, almost as important, could mean the difference between being readmitted at a later date and staying on the road to recovery.
The research team includes Mortazavi along with researchers and collaborators from Yale University and the American College of Cardiology. The team’s work was recently published in JAMA Network Open.
“One of the reasons that the medical field has suddenly become very open to working with computer scientists is that, under certain healthcare legislature, there’s a lot of clauses, now, regarding heart failure, readmissions and quality of care improvements,” said Mortazavi. “Sometimes patients are even kept in the hospital longer post-procedure if it means that they don’t come back later.
“We are using the newest, latest and greatest computer science, artificial intelligence and machine learning techniques to see how we improve patient care, clinical decision making and the prediction of what’s going to happen,” said Mortazavi.
The challenge, he explained, is that many of these techniques were developed on the computer science side, where decision-making parameters are easily defined and little mistakes – such as a Fitbit miscounting a few steps – is not a big deal. However, for health care purposes, results and conclusions must not only be highly accurate, but also need to be interpretable by whoever is using them.
To achieve this interpretability, the team looked at the categorization of high- and low-risk patients as a spectrum, rather than a black-and-white decision – shifting away from traditional pattern recognition techniques that match a data set to a specific, programmed conclusion.
As Harlan Krumholz, a cardiologist and professor at Yale University, explained in a recent news story, the team is discovering that machine learning has the potential to improve the ability to predict risk better than traditional approaches. In the future, such techniques could enable doctors and engineers to personalize risk estimates to a much greater extent.
