The search described in The nature of biomedical engineering, found that the model was more effective at identifying problems such as pneumonia, collapsed lungs, and lesions than other self-supervised AI models. In fact, they were comparable in accuracy to human radiologists.
While others have tried to use unstructured medical data in this way, this is the first time the team’s AI model has learned from unstructured text and the performance of matched radiologists, and has demonstrated the ability to predict multiple diseases from a given x-ray using a high degree of accuracy, he says. Ikin Teo, an undergraduate at Stanford University and a visiting scholar who co-authored the report.
“We’re the first to do that and we’re demonstrating it effectively in the field,” he says.
The model’s code has been made publicly available to other researchers with the hope that it will be applied to CT scans, MRIs and echocardiography to help detect a wide range of diseases in other parts of the body, says Pranav Rajpurkar, MD, assistant professor of biomedical informatics at the Blavatnik Institute at College of Medicine. Harvard Medicine, who led the project.
“We hope that people can apply this out of the box to other chest x-ray data sets and the diseases they care about,” he says.
Rajpurkar is also optimistic that diagnostic AI models that require minimal supervision can help increase access to health care in countries and societies where specialists are scarce.
“It makes a lot of sense to use the richer training signal than reports,” says Christian Liebig, director of machine learning at German startup Vara. Artificial intelligence is used to detect breast cancer. “It’s a huge achievement to reach this level of performance.”