A new artificial intelligence computer program created by researchers at the University of Florida and NVIDIA can generate doctors’ notes so well that two physicians couldn’t tell the difference, according to an early study from both groups.
In this proof-of-concept study, physicians reviewed patient notes — some written by actual medical doctors while others were created by the new AI program — and the physicians identified the correct author only 49% of the time.
A team of 19 researchers from NVIDIA and the University of Florida said their findings, published Nov. 16 in the Nature journal npj Digital Medicine, open the door for AI to support health care workers with groundbreaking efficiencies.
The researchers trained supercomputers to generate medical records based on a new model, GatorTronGPT, that functions similarly to ChatGPT. The free versions of GatorTron™ models have more than 430,000 downloads from Hugging Face, an open-source AI website. GatorTron™ models are the site’s only models available for clinical research, according to the article’s lead author Yonghui Wu, Ph.D., from the UF College of Medicine’s department of health outcomes and biomedical informatics.
“In health care, everyone is talking about these models. GatorTron™ and GatorTronGPT are unique AI models that can power many aspects of medical research and health care. Yet, they require massive data and extensive computing power to build. We are grateful to have this supercomputer, HiPerGator, from NVIDIA to explore the potential of AI in health care,” Wu said.
UF alumnus and NVIDIA co-founder Chris Malachowsky is the namesake of UF’s new Malachowsky Hall for Data Science & Information Technology. A public-private partnership between UF and NVIDIA helped to fund this $150 million structure. In 2021, UF upgraded its HiPerGator supercomputer to elite status with a multimillion-dollar infrastructure package from NVIDIA, the first at a university.
For this research, Wu and his colleagues developed a large language model that allows computers to mimic natural human language. These models work well with standard writing or conversations, but medical records bring additional hurdles, such as needing to protect patients’ privacy and being highly technical. Digital medical records cannot be Googled or shared on Wikipedia.
To overcome these obstacles, the researchers stripped UF Health medical records of identifying information from 2 million patients while keeping 82 billion useful medical words. Combining this set with another dataset of 195 billion words, they trained the GatorTronGPT model to analyze the medical data with GPT-3 architecture, or Generative Pre-trained Transformer, a form of neural network architecture. That allowed GatorTronGPT to write clinical text similar to medical doctors’ notes.
Read more here.