Currently, doctors have many ways to predict the health of a patient. None of them, however, is universal, and many pathologies (for example, heart attacks) are very difficult to predict. Scientists have demonstrated that computers capable of self-learning can perform even better than standard medical practices and significantly improve the quality of prediction. If this practice is implemented, the new method will help save thousands, if not millions of lives every year.
Every year, about 20 million people die from cardiovascular diseases, including heart attacks, strokes, clogged arteries and other cardiovascular diseases. In order to try to predict such complications, doctors in Western countries use the guidelines of the American College of Cardiology / American Heart Association (ACC / AHA). They are based on eight risk factors, including age, blood cholesterol levels and blood pressure, from which the doctor tries to compose a single picture of the disease.
For many cases, this approach is often overly simplistic, in addition, other factors can affect the patient's body, as a result of which cardiovascular diseases can develop. In a new study, Stephen Wan, an epidemiologist at the University of Nottingham in the UK, compared the ACC / AHA directives to four machine learning algorithms: random forest, logistic regression, gradient boosting, and neural network. All four algorithms were aimed at analyzing a lot of data that, in theory, would allow AI to make medical predictions better than humans. In this case, the data was obtained from electronic health records of 378,256 patients in the UK. The goal was to find sample recordings that were associated with cardiovascular events.
First, artificial intelligence (AI) algorithms had to train on their own. They used about 78% of the data - roughly 295,267 records - to search for patterns and create their own internal “recommendations”. Then they tested themselves on the rest of the documents. Using data from 2005, the algorithms predicted which patients would get heart and vascular problems over the next 10 years, and then tested their assumptions using 2015 records. Contrary to the ACC / AHA guidelines, machine learning was allowed to take into account 22 more data points, including ethnicity, arthritis, and kidney disease.
As a result, all four AI methods were found to be much more efficient at forecasting than the ACC / AHA recommendations. Using AUC statistics (where 1.0 is 100% accurate), the ACC / AHA directives have reached 0.728. The four new methods ranged from 0.745 to 0.764, as Wen's team reported in PLOS ONE magazine. In the test sample, about 83,000 entries took part, and in the battle between AI and man, the machines "saved" 355 more patients. This is because, Wen says, prediction often leads to prevention, through cholesterol lowering or dietary changes.
Some of the risk factors that machine learning algorithms have identified as the strongest predictors are not included in the ACC / AHA guidelines. These include, for example, severe mental illness and oral administration of corticosteroids. Meanwhile, none of the parameters that are on the ACC / AHA list are among the 10 most important predictors by machine (and even diabetes). In the future, Weng hopes to include other social and genetic ones to further improve the accuracy of the algorithms.
Vasily Makarov