AI tools may fail during major medical diagnosis – researchers

At the first moment when it comes to the role of artificial intelligence in terms of critical health data, the US team of researchers has said that AI in the medical room should carefully test performance in different populations, as deep learning models may be short.

This result should be paused by those who are considering the rapid deployment of AI platforms, which is not exactly limited to their performance in real-world clinical conditions, reflecting their placement, a team from the Icahn School of Medicine at Mount Sinai School of Medicine observed.

According to a study published in the special edition PLOS Medicine on machine translation and healthcare, XI-XI diagnostic tools used to determine X-ray imprints of pneumonia were significantly reduced when performing tests based on data from external health systems.

These findings suggest that deep learning patterns may not work as accurately as expected.

"Advanced in-depth training for medical diagnostics can be broadly expressed, but it can not be taken for granted as patient population and imaging techniques vary widely from one institution to another," said senior author Eric Oermann, a neurosurgeon engineer at Icahn Medical School in Sinai.

To achieve this balance, researchers estimated the AI ​​patterns identified pneumonia of 158,000 chest X-rays in three medical facilities – the National Institutes of Health, Mount Sinai Hospital, and the University of Indiana Hospital.

In three of the five comparisons, the performance of the CNNs in diagnosing X-rays from hospitals located outside its network was significantly lower than the X-rays in the original health system.

However, CNN was able to detect a hospital system in which the X-ray was obtained with a high degree of precision and pondered their prediction task based on the spread of pneumonia in the institution.

"If AI systems are used for medical diagnosis, they should be suitable for careful consideration of clinical issues, testing in different real-time scenarios, and carefully scrutinized to determine how they affect the exact diagnosis," explains the study's first author, John Seek.