NYU Langone's Lang1, trained on 10 million patients' records, beats larger AI models at predicting hospital outcomes
The Nature Medicine study reports gains of up to 24 percent on insurance-denial prediction but only 1.6 percent on in-hospital mortality versus models such as DeepSeek R1.
Why it matters
The study is evidence that a smaller model trained on a health system's own records can beat much larger general models on routine operational predictions, which matters for hospitals weighing what AI to deploy. Lang1 builds on NYU's earlier NYUTron, is nearly 70 times larger and was trained on 80 billion tokens of clinical notes mixed with 627 billion tokens of general text. The gains are largest where internet text is thin, such as insurance denials, and smallest for mortality. NYU Langone plans to offer Lang1 and its ReMedE benchmark to other hospitals; watch for results beyond the Beth Israel validation.
Key facts
- Trained on records of more than 10 million patients over 20 years
- 80 billion clinical-note tokens plus 627 billion general-text tokens
- Nearly 24% better than DeepSeek R1 at predicting insurance denials
- 9% better at 30-day readmissions, 1.6% at in-hospital mortality
- Nearly 70 times larger than NYUTron
- Validated at Beth Israel in Boston