AI generates medical notes indistinguishable from human doctors
It could free doctors from the boring but critically important task of writing notes for patients.
In a remarkable advancement for artificial intelligence, researchers from the University of Florida (UF) and NVIDIA have developed a new AI computer program capable of generating doctors’ notes indistinguishable from those written by actual physicians.
The proof-of-concept study revealed a striking result: even seasoned doctors correctly identified the author of the notes — AI or human — only about half the time. They might have just flipped a coin. This underscores the AI’s sophisticated ability to mimic professional medical writing.
Paging Dr. AI
One of these two notes was written by a human doctor, the other by AI. Can you tell which is which? The answer is revealed at the end of this article.
At the heart of this innovation is GatorTronGPT, a model akin to ChatGPT, created by a team of 19 researchers from NVIDIA and the University of Florida. This model, along with its predecessor GatorTron™, has become a cornerstone in medical research and healthcare, with over 430,000 downloads from the open-source AI website, Hugging Face.
Yonghui Wu from the UF College of Medicine highlights the uniqueness of these models, which require extensive data and computing power. This is where the university’s HiPerGator supercomputer comes in, built with a $150 million investment in 2021.
Creating a language model that accurately replicates medical notes involves unique challenges, such as ensuring patient privacy and managing complex medical terminology. To address this, the team meticulously processed UF Health’s medical records, carefully anonymizing data from 2 million patients and preserving a vast corpus of 82 billion medical words. This dataset, combined with an additional 195 billion words, ensured that GatorTronGPT was trained with an extensive clinical vocabulary and range of language patterns.
The GatorTronGPT model is first and foremost a proof of concept, the brainchild of fundamental research in computer science. But that doesn’t mean it can’t have practical applications. On the contrary, GatorTronGPT can alleviate the burdensome task of medical documentation, freeing up resources and allowing doctors to focus more on patient care. Efforts are already underway at UF’s innovation center to bring a commercial version of this AI tool to fruition.
The findings appeared in the journal Digital Medicine.
The first note was written by a human, the second note was made by AI.
