NYU Langone's Lang1 beats much larger AI models at predicting hospital outcomes

NYU Langone Health researchers described Lang1, a family of AI models that reads doctors' clinical notes to predict what happens to hospital patients next, in a release on a study published in Nature Medicine on 9 November.

Lang1 was trained on records from more than 10 million patients over 20 years of care: 80 billion tokens from electronic health record notes, mixed with 627 billion tokens of general text such as books, papers and Wikipedia. It predicts in-hospital decline and mortality, 30-day readmissions, length of stay and insurance denials.

Against general models including GPT-4o, Llama and DeepSeek R1, which the authors say is 671 times Lang1's size, it was nearly 24 percent better at predicting insurance denials and 9 percent better at predicting readmissions. The gap shrank to 1.6 percent for in-hospital mortality, a topic the authors note is common in internet text that general models learn from.

The team also built ReMedE, a test that checks a model's accuracy on five real hospital prediction tasks before deployment, and validated the results at Beth Israel in Boston. The authors estimate that even modest gains on readmissions and denials could save a hospital close to $4 million a year. NYU Langone plans to offer Lang1 and ReMedE to other hospitals as a confidential service; care decisions stay with clinicians.