+1 386-243-9402 MON – FRI : 09:00 AM – 05:00 PM

What is a Large Language Model?

A large language model (LLM) is a type of artificial intelligence (AI) model which is fed with a massive dataset of information from the Internet. A set of complex computer algorithms analyzes the information and learns the statistical relationships between words and phrases. This process is called training. After the LLM is trained, it can generate new information similar to the information it was trained on. LLMs are at their early stages but are being developed by many organizations. The LLMs have the potential to revolutionize the way we interact with computers. LLMs are already being used to create more natural and engaging user interfaces like Bard and chatGPT. LLMs can also improve the accuracy of machine translation and develop new ways of generating creative content.

I believe one of the biggest areas of opportunity to take advantage of LLMs is healthcare. Researchers at NYU Langone collaborated with chipmaker Nvidia to develop an LLM which can predict a patient’s risk of readmittance within 30 days. Readmission is one of the costliest episodes in treatment in the country, racking up $41.3 billion in medical costs annually. The U.S. government imposes financial penalties on hospitals with a high rate of readmissions. It is not just a financial matter. Patient safety, quality care, and quality of life are major issues for patients, and readmissions can interfere will all those issues.  

The NYU AI tool is called NYUTron. The name is weird but the massive data the researchers have amassed is impressive. They used 10 years of health records at NYU Langone Health to train NYUTron with 4 billion words of clinical notes representing 400,000 patients. The model, which the researchers believe is quite accurate, is designed to identify patients in need of clinical intervention. The researchers apply four predictive algorithms to the LLM to predict, in addition to likely readmission, the length of a patient’s hospital stay, the likelihood of mortality, and the chances of insurance claims being denied.

NYUTron has been rolled out in the NYU health system’s six inpatient facilities and applied to over 50,000 patients. Physicians receive email notifications of patient readmission risk or other risks. The idea of predicting readmissions is not new, but building an LLM based on all the structured data in electronic health records may capture insights which have not been considered before.

Building an LLM of this scale is not something most hospitals can afford to do. However, they may be able to piggyback on what NYU has done. Dr. Oermann at NYU said, “Not all hospitals have the resources to train a large language model from scratch in-house, but they can adopt a pretrained model like NYUTron and then fine-tune it with a small sample of local data using GPUs in the cloud,” Oermann said. “That’s within reach of almost everyone in healthcare.”

We are entering a period when breakthroughs can be expected from the application of AI. Ethical and integrity issues abound and must be taken seriously. Humans need to stay in the loop. I believe advances in healthcare over the next ten years will surpass advances of the past 100 years.

For information about NYUTron and the latest generative AI efforts in healthcare, see Fierce Healthcare.