Working with 10 medical experts and using more than 100 patient discharge summaries, researchers at IBM and MIT compiled more than 2,000 questions and 1,000 triggers of questions that physicians may ask when looking at a patient’s EHR. They categorized each question or trigger written by the medical experts into groups such as symptom/sign, demographics and test results, making it easier for physicians to navigate through the questions.
The team then trained a data model to do this organically. For instance, if an EHR notes that a patient had a mass in their neck, one question generated may be about its size or color, in the category of symptoms.
They found that their model produced a high-quality question 62.5 percent of the time in response to a prompt, but only if given more context. Without the context, the question generation quality declined.
“Our results demonstrate that existing machine learning systems, including large-scale neural networks, struggle with the tasks we propose. We encourage the community to improve on our baseline models,” the researchers wrote. They also opened their machine learning tool to the public for continued work.
At the Becker's 11th Annual IT + Revenue Cycle Conference: The Future of AI & Digital Health, taking place September 14–17 in Chicago, healthcare executives and digital leaders from across the country will come together to explore how AI, interoperability, cybersecurity, and revenue cycle innovation are transforming care delivery, strengthening financial performance, and driving the next era of digital health. Apply for complimentary registration now.