NYU’s specialized AI beats general models on hospital predictions, study finds

NYU Langone Health is pitching its new hospital AI as a way to improve daily care decisions. The evidence announced November 9 establishes a narrower result: its Lang1 models predicted hospital outcomes better than several general-purpose rivals. The biggest reported advantage concerned insurance denials; the improvement in predicting deaths during hospital stays was much smaller.

The study, published online in Nature Medicine, describes Lang1’s development, testing and deployment in real-world settings. It is a family of prediction models trained on records from more than 10 million patients across 20 years of care, alongside a large body of general text.

The advantage depends on the hospital task

Researchers compared Lang1 with models including ChatGPT-4o, Meta’s Llama and DeepSeek R1. According to NYU Langone’s account of the study, Lang1 was nearly 24% better than DeepSeek R1 at predicting insurance denials. The announcement also reports improvements of 9% for 30-day readmissions and 1.6% for in-hospital mortality.

Those figures describe gains in a statistical measure called area under the receiver operating characteristic curve, or AUROC. It measures how well a model separates groups—for example, patients who will return to the hospital from those who will not. They are not measured reductions in deaths, readmissions or denied claims.

Nearly 24%Insurance denials

NYU Langone reports the largest improvement against DeepSeek R1 on this prediction task.

9%30-day readmissions

The reported gain concerns prediction performance, not fewer return hospital visits.

1.6%In-hospital mortality

The mortality prediction gain was smaller than the other two reported improvements.

The researchers offer an explanation for that gap. Language about mortality appears frequently online, giving general-purpose models more relevant material to learn from. Insurance-denial information is less represented there, they argue, making hospital records especially valuable for that task. That is the authors’ explanation, rather than a separate finding about patient care.

Learning what doctors’ notes leave unstated

Lang1’s training combined 80 billion tokens from electronic health record notes with 627 billion tokens of general text, including books, Wikipedia and academic papers. Tokens are the pieces of words AI models process. The researchers say adding general text helped Lang1 handle the messy language in clinical notes without hurting its clinical prediction performance.

A note can describe symptoms and treatment without spelling out what will happen next. To teach that connection, the team used supervised fine-tuning: showing the model millions of notes and comparing its predictions with patients’ actual outcomes. The task was to infer future events, not simply repeat information already written in a record.

The family builds on NYUTron, the team’s earlier clinical-note prediction model. NYU Langone says Lang1 is nearly 70 times larger and was trained on more than 10 times as much material. The study therefore combines specialized records with broader language training, rather than relying on hospital text alone.

Turning risk scores into care decisions

Lang1 is designed to support human care providers, not replace their judgment. NYU Langone describes two ways its predictions could inform care:

• A high predicted risk of death could prompt closer monitoring and palliative care consultations.
• A high predicted readmission risk could prompt better discharge planning or consideration of a longer hospital stay.

The team also developed ReMedE, an evaluation tool that checks prediction accuracy across five tasks using operational hospital outcomes. It is intended to expose gaps before deployment. Researchers validated their results at Boston’s Beth Israel as well as NYU Langone, providing a test beyond the health system that developed the models.

NYU Langone plans to offer Lang1 and ReMedE to other hospitals as a confidential service, seeking further evidence that the system can work nationally. The authors also estimate annual savings could approach $4 million per hospital if model-guided decisions produce modest improvements. That figure is a projection based on costs drawn from existing literature—not savings demonstrated by the study.