An artificial intelligence algorithm has bested experts at predicting patient deaths from heart disease. In a study published recently in the journal PLOS One, researchers from the Francis Crick Institute make yet another case for using A.I. to inform medical diagnoses.
“We’ve shown that you can give a computer someone’s medical records … and predict how likely a patient with heart disease is to die,” Andrew Steele, a Crick researcher and first author of the paper, told Digital Trends. “Traditional models get experts to select the most relevant variables for making these kinds of predictions, but we did just as well without telling the computer … the most important or relevant things to take into account.”
In their study, Steele and his team worked with researchers from the Farr Institute of Health Informatics Research and University College London Hospitals NHS Foundation Trust to test whether a self-taught model could outperform experts at predicting deaths from coronary artery disease.
The model created by Steel and his colleagues was compared to expert predictions, which take into account some 27 variables, including age, gender, and physical ailments. The Crick algorithm was tasked with finding patterns and useful variables from list of 600. With access to 80,000 anonymized patient health records, the A.I. was able to outperform medical experts and identified new variables that doctors had overlooked.
“They say that making predictions is hard, especially about the future, and building these models can be difficult and time-consuming,” Steele said. “The great advantage of A.I. is that, done right, you can just throw all the data in and let the computer work out what’s relevant, which could save future researchers a lot of time.”
To be sure, the A.I. performed just fractionally better than its human counterparts. Given two randomly selected patients, the algorithm could predict which patient would die first just one percent more often than medical experts, Steele said. So we aren’t exactly talking about a breakthrough.
We are, however, talking about small steps toward more accurate and effective diagnoses. As we’ve written about before, A.I. has the potential to revolutionize healthcare by supporting physicians with the more meticulous and data-driven part of the job.
“I think at first these systems are going to be assisting rather than replacing doctors,” Steele said. “Doctors already use tools to, for example, check your risk of a heart attack in the next few years before prescribing certain drugs. A.I. will help us develop more of these tools for different conditions, and help doctors and patients make better decisions. In the longer term, I think we’ll see A.I. systems making recommendations for how to treat a patient, and we’ve already got computers interpreting things like scan results. But, for now at least, a human doctor is very important in understanding the output of these models, and helping patients make decisions based on them.”