Language & LLMs
A large clinical language model improved performance on five electronic health record language tasks

In a 2022 paper in npj Digital Medicine, researchers from the University of Florida and Nvidia report that a clinical language model called GatorTron improved performance on five tasks that help computers read electronic health records. The largest GatorTron model had 8.9 billion parameters. Parameters are the adjustable parts of a model that learn from training data. The model was trained on more than 90 billion words of text, including more than 82 billion words of de-identified clinical text. The researchers report a 9.6% improvement in accuracy on natural language inference, the task of deciding whether one statement follows from another, and a 9.5% improvement on medical question answering.
What did the study test?
The team evaluated GatorTron on five clinical language tasks. These included clinical concept extraction, medical relation extraction, semantic textual similarity, natural language inference, and medical question answering. The study also examined how increasing the number of parameters and increasing the size of the training data affected performance on those tasks.
How much did GatorTron improve?
The researchers scaled a clinical language model from 110 million to 8.9 billion parameters. The abstract reports a 9.6% improvement in accuracy on natural language inference and a 9.5% improvement on medical question answering. Those two tasks were among the five tested.
Why do these tasks matter?
These tasks are part of natural language processing, which lets computers interpret human language in text. Clinical concept extraction finds medical terms in notes, and medical relation extraction finds connections between those terms. Medical question answering can help systems respond to questions about care. The authors say the models could be applied to medical AI systems to improve healthcare delivery.
Frequently asked questions
How large is GatorTron?
The abstract reports GatorTron models scale from 110 million to 8.9 billion parameters.
What data did it train on?
GatorTron was trained from scratch on more than 90 billion words of text. More than 82 billion of those words were de-identified clinical text.
Where can the models be found?
The abstract says the GatorTron models are publicly available through Nvidia's NGC catalog.
Limits and what's next
The abstract describes an evaluation on five clinical language tasks using de-identified clinical text and other text. It does not report a test inside a hospital or clinic, and it does not report whether the gains changed patient outcomes. This is one study, and the abstract reports no independent replication in another health system. The authors say the models could be applied to medical AI systems to improve healthcare delivery, but the abstract does not report those downstream results.
Source: Xi Yang et al., “A large language model for electronic health records,” npj Digital Medicine, 2022.