AI learns to predict heart attacks better than doctors do
A set of algorithms, developed by researchers at the University of Nottingham in the United Kingdom, can predict which patients are most at risk of a heart attack or stroke at over 75 percent accuracy. Although you may feel like this number should be even higher, it is already nearly 8 percent above the average accuracy level of highly trained doctors.
In the United States alone, someone has a heart attack every 34 seconds. Globally, 20 million people die each year as a result
of heart attacks, strokes, or other cardiovascular or circulatory malfunctions. Although doctors are capable of determining which individuals are at risk of developing heart disease and other cardiovascular problems, determining when an individual may suffer from one is much more difficult.
An epidemiologist at the University of Nottingham, Stephen Weng explained why it is often challenging for doctors to make accurate predictions. The human body is an intricate, complex, ever- developing system with a great deal of counterintuitive functions. For example, doctors use high body fat content as a red ag indicating that the individual may be at risk of heart disease; how- ever, in many cases, body fat becomes impermeable and can protect the heart against disease. Although it is dif cult for a doctor to make these sort of predictions, Weng explained that the new AI algorithms can train themselves to explore and gain understanding about these nuances.
These artificial intelligence algorithms are self-taught and continually using new data to make faster, smarter decisions without human instruction. Using nearly 300,000 available patient records from 2005, the algorithms create guidelines for themselves to predict which patients would have a heart attack or stroke within the next 10 years.
These AI algorithms are self-taught and continually using new data to make faster, smarter decisions without human instruction. Using nearly 300,000 available patient records from 2005, the algorithms create guide- lines for themselves to predict which patients would have a heart attack or stroke within the next 10 years.
The algorithm began by using the rst 80 percent of the records as input data to make an initial rule book of sorts. When making the rule book, the algorithms considered metrics that doctors commonly use — age, weight, body fat content — in addition to categories the researchers thought may be interesting to examine, such as arthritis, kidney disease, or mental illness. The AI then tested its new rule book on the remaining data, improving the rules and making adjustments to optimize it’s prediction.
In addition to the AI making predictions with 8 percent more accuracy than doctors, it also made 1.6 percent fewer false predictions. If the algorithm had existed at the time, it could have saved 355 deceased patients whose records it utilized during testing. These numbers are good, but researchers are determined to make the algorithm even better.
These machine-learning algorithms are dif cult to adjust, however. Because the algorithm is self- taught, it is dif cult for the machine’s programmers to fully understand of all of its internal functionality. The programmers are hesitant to make large adjustments, fear- ing that they may undo some of the learning and reverse some of the progress the machine has already made for itself.
The researchers at Nottingham worry that some doctors, who take pride in their hard-earned expertise, may be hesitant to adapt these machine-learning methods; however, if further testing continues and the AI is widely implemented in hospitals, the algorithm could start saving hundreds of thousands of lives every year.