Illustration: Aïda Amer/Axios
A new review concludes that the claims that artificial intelligence is better than human experts at interpreting medical scans and images are likely overblown.
Why it matters: There's legitimate hope that AI could improve patient care and cut costs. But there's also a real risk that people could be hurt if biased or mistake-prone algorithms are rushed into medicine.
Background: One of the first commercial uses for AI in health care has been the work of interpreting medical imagery such as X-rays or CT scans. Deep learning algorithms can be trained on massive sets of medical images, and then evaluate scans faster and potentially better than a human. Or at least that's the hope.
In a review published on March 25 in the British Medical Journal, a team of researchers looked at more than 80 studies over the last 10 years that compared the performance of deep learning algorithms on medical imaging to expert clinicians.
- They were unimpressed by the studies: Just two were randomized clinical studies — the gold standard for science — and they found that more than two-thirds of reviewed studies were at high risk of bias.
- Despite those methodology problems, some three-quarters of the studies purported to find that AI was comparable to or better than human clinicians at the work of interpreting medical imaging, which helps fuel the hype.
The bottom line: Just as doctors take the Hippocratic Oath, we need to ensure that AI will do no harm before it becomes an integral part of medicine.
Go deeper: The two uncertainties of the coronavirus