[ad_1]
“The epidemic was a major test for AI and medicine,” says Driggs, who also uses a machine learning tool to help doctors with the epidemic. “It would be a long way for people to be on our side,” he says. “But I don’t think we passed that test.”
Both groups found that researchers repeated the same errors in the way they taught or tested their weapons. Misconceptions about data type often mean that trained types do not work as they claim.
Wynants and Driggs believe AI has a chance to help. But they are concerned that it could be harmful if it was built incorrectly because it could miss the disease or reduce the risk of at-risk patients. “There are a lot of attractions in the field of machine learning and what they can do today,” says Driggs.
Unfulfilled expectations encourage the use of these tools before they are ready. Wynants and Driggs both claim that several of the algorithms they looked at were already used in hospitals, and some were sold by manufacturers themselves. “I’m afraid he might injure patients,” Wynants said.
So what’s wrong? And how do we bridge these differences? If there are other problems, then the epidemic has explained to many experts that the way AI tools are made should change. “The epidemic has exposed the problems we have been discussing for some time,” Wynants said.
What went wrong
Many of the problems identified are linked to the unpredictable nature of what the researchers used to create their weapons. Information about covid patients, including medical examinations, was collected and shared during the global epidemic, often by doctors who were trying to cure those patients. The researchers want to help urgently, and these are the only ones who contribute in public. But this meant that many weapons were built using information or false information from unknown sources.
Driggs points out the problem of so-called Frankenstein data, which is compiled from a number of sources and can be computed. This means that some tools end up being tested on the same type of training they were taught, making them look more accurate than they really are.
It also corrupts other sources of information. This could mean that researchers lack the resources that make their studies worthwhile. Many unknowingly used a set that contained the breasts of children who did not have covid as their examples of how covid-free cases appear. As a result, AI learned to recognize children, not covids.
The Driggs team taught their brand using a set of data that included mixed recipes that patients slept and stood up for. Because patients who are diagnosed with sleep apnea can become very ill, AI has mistakenly learned to predict human risk.
In some cases, some AIs are found to be based on the records that other hospitals use for testing. As a result, letters from hospitals with more serious cases predicted covid risk.
Errors like these seem obvious. It can be redesigned and modified colors, if researchers are aware. It is possible to acknowledge the error and produce an incomplete, but misleading type.
But most of the tools were developed by AI researchers who had no medical expertise in diagnosing errors or by medical researchers who had no mathematical ability to correct the errors. A well-known problem that Driggs points out is the inclusion, or bias, that is defined when it is stored.
For example, many medical tests were performed based on whether the radiologists who developed them said they were showing covid. But this includes, or does include, any doctor’s arguments in the actual facts of the record. It would be better to focus on clinical analysis and results of PCR tests than to focus on a single physician, says Driggs. But there is not always enough time to study in hospitals.
This did not stop some of these weapons from being rushed to hospitals. Wynants says it is not known which ones are being used or how. Hospitals sometimes report using a diagnostic tool, making it difficult to see how many doctors they rely on. “There are a lot of secrets,” he says.
Wynants asked one company that was promoting in-depth learning strategies to share more of its approach but they did not hear. He later obtained several printed copies from investigators detained by the company, all of whom are at high risk. “We don’t really know what the company did,” he says.
According to Wynants, some hospitals sign anonymous agreements with medical AI vendors. When she asks the doctors about the program or programs she is using, they sometimes tell her that they are not allowed to talk.
How to prepare
What is it planning? A lot can be helpful, but when problems arise it is a big question. It is very important that we make good use of the dates we have. The simple fact is that AI teams are more closely associated with physicians, says Driggs. Researchers also need to share their colors and demonstrate how they were trained so that others can practice and build. “These are two things we can do today,” he says. “And that’s about 50% of the problems we’ve found.”
Access to data would be easier if the species were identical, says Bilal Mateen, a leading medical specialist at Wellcome Trust, a global health research organization based in London.
[ad_2]
Source link
