[ad_1]
This is because health data such as medical imaging, vital signs, and wearable device data can change for reasons unrelated to specific health conditions, such as lifestyle or background noise.Machine learning algorithms pervasive in the tech industry are very good at finding patterns, and they can Shortcuts to discovering the “correct” answer This doesn’t work in the real world. Smaller datasets make it easier for algorithms to cheat in this way and create blind spots that lead to poor clinical outcomes. “Community fools [itself] The idea that we’re developing models that work better than they actually are,” Berisha said. “It’s driving the AI hype even further. “
The problem has led to a striking and worrying pattern in certain areas of AI-powered healthcare research, Berisha said. In studies using algorithms to detect signs of Alzheimer’s or cognitive impairment in speech recordings, Berisha and his colleagues found that large studies reported less accuracy than smaller ones — in contrast to what big data should provide on the contrary.One review research attempts to draw from medical scans and other Similar patterns have been reported for studies attempting to detect autism through machine learning.
The dangers of algorithms that work well in preliminary studies but perform differently on real patient data are not hypothetical. A 2019 study finds that a system used for millions of patients prioritizes additional care for those with complex health problems Putting white patients before black patients.
Avoiding such biased systems requires large, balanced datasets and careful testing, but skewed datasets are the norm in health AI research due to historical and persistent health inequalities.One 2020 study by Stanford researchers Discover 71% of the data used in applied research deep learning U.S. medical data comes from California, Massachusetts or New York, and the other 47 states have little or no representation. Low-income countries are barely represented in AI healthcare research.review published last year More than 150 studies using machine learning to predict diagnosis or disease course concluded that most “showed poor methodological quality and were at high risk of bias”.
Two researchers concerned with these shortcomings recently launched a project called Nightingale Open Science Try to improve the quality and size of datasets available to researchers. It works with health systems to collect medical images and related data from patient records, anonymize them, and make them available for nonprofit research.
Nightingale co-founder and associate professor Ziad Obermeier at UC Berkeley hopes that providing access to this data will encourage competition, leading to better outcomes, similar to a large-scale, open collection of images Help drive progress in machine learning. “The core of the problem is that researchers can do whatever they want with health data because no one can check their results,” he said. “data [is] lock. “
Nightingale joins other projects trying to improve AI in healthcare by increasing data access and quality.This Gap Fund Support the creation of machine learning datasets representative of low- and middle-income countries and work on healthcare; a new project Birmingham University Hospital in the UK, with support from the National Health Service and the Massachusetts Institute of Technology, is developing criteria to assess whether artificial intelligence systems are anchored in unbiased data.
Martin, editor of the UK’s Epidemic Algorithms Report, is a fan of such AI-specific projects, but says the future of AI in healthcare also depends on health systems modernizing it. Often crumbling IT infrastructure. “You have to invest in the root of the problem to see the benefits,” Mateen said.
More great Wired stories
[ad_2]
Source link
