Researchers doubt the accuracy of AI’s medical diagnosis

AI models ignored clinically significant indicators and instead relied on features such as text markers or patient positioning

New York: Artificial intelligence (AI) models like humans tend to look for shortcuts. In the case of AI-assisted disease detection, these shortcuts could lead to misdiagnosis if deployed in clinical settings, the researchers warn.

A team from the University of Washington in the United States examined several models recently proposed as potential tools to accurately detect Covid-19 from chest x-rays, also known as chest x-rays.

The results, published in the journal Nature Machine Intelligence, showed that instead of learning a true medical pathology, these models instead rely on shortened learning to establish false associations between medically irrelevant factors and the condition of the patient. sickness.

As a result, the models ignored clinically significant indicators and instead relied on features like text markers or patient positioning that were specific to each data set to predict whether someone had Covid-19.

“A doctor would generally expect a discovery of Covid-19 from an x-ray to be based on specific patterns in the image that reflect disease processes,” said lead co-author Alex DeGrave of the UW Medical Scientist Training Program.

“But rather than relying on these models, a system using shortened learning could, for example, judge that a person is old and thus infer that they are more likely to have the disease because it is more common. in older patients.

“The shortcut is not inherently bad, but the association is unexpected and not transparent. And that could lead to an inappropriate diagnosis,” DeGrave said.

“Less robust”

Shortcut learning is less robust than true medical pathology and generally means the model will not generalize well outside the original setting, the researchers said.

The combination of the lack of robustness with the opacity typical of AI decision making can make these AI models subject to a condition known as “worst case confusion”, due to the lack of data from training available for such a new disease.

This scenario increased the likelihood that the models relied on shortcuts rather than learning the underlying pathology of the disease from the training data, the researchers noted.

For all the latest News, opinions and views, To download application.

To select Tongue To read in Urdu, Hindi, Marathi or Arabic.