In the United States, over 600,000 people die every year from heart attacks. Turns out, heart disease is the leading cause of death for both men and women. Thanks to a team of researchers from the University of Nottingham, a machine-learning algorithm has been developed that can predict a person’s likelihood of having a heart attack or stroke as well as any doctor.
THE ROBOT DOCTOR -
Traditional guidelines for estimating a patient’s cardiovascular risk were created by the American College of Cardiology and American Heart Association. These guidelines are based on factors such as age, cholesterol level, and blood pressure. Using the ACC and AHA guidelines the current system predicts the occurrence of a heart attack with 72.8 percent accuracy.
Challenging that statistic is Stephen Weng and his team of researchers from the University of Nottingham. The researchers compared the accuracy of the ACC/AHA guidelines by building four machine learning algorithms.
Each algorithm – random forest, logistic regression, gradient boosting, and neural networks – analyzed the data from 378,256 electronic patient records. The machine’s goal was to find patterns within the records that are related to cardiovascular events.
In this case, the AI algorithms had to train themselves. The algorithms used about 78% of the data to search for patterns – think of this as the machine’s foundation for understanding the data. The algorithms then tested themselves on the remaining data and tested record data from 2005 to predict which patients would have their first heart attack or cardiovascular event over the next 10 years. Those predictions were then checked against 2015 records. All four AI algorithms outperformed the traditional ACC and AHA guidelines with statistics ranging from 74.5 to 76.4 percent.
IMPLICATIONS - According to Science, Weng’s test sample of 83,000 patient records could have saved about 355 additional patients. Weng believes prediction often leads to prevention through cholesterol-lowering medication and changes in diet.
The algorithms identified several risk factors, like severe mental illness and taking oral corticosteroids, as the strongest predictors for cardiovascular events. The AHA and AC guidelines do not include these predictors. To counter, none of the algorithms accounted for diabetes which the ACC/AHA guidelines consider to be a top 10 predictor for heart attacks. Stephen Weng and his team plan to include more lifestyle and genetic factors to improve the algorithm’s accuracy in the future.