TL;DR:
- NYU Langone Health and NVIDIA collaborated to develop NYUTron, a large language model (LLM) for predicting patient readmission.
- NYUTron provides AI-driven insights to identify patients at risk of readmission, enabling clinical interventions.
- The model has been deployed in NYU’s healthcare system, demonstrating promising results in reducing readmission rates.
- NYUTron was trained on extensive electronic health records, achieving significant accuracy improvements over traditional models.
- The collaboration also resulted in predictive algorithms for hospital stay length, in-hospital mortality, and insurance claim denials.
Main AI News:
Patient discharge marks a crucial milestone in their journey toward recovery. However, for a significant portion of hospital patients in the United States, it’s not the end of the road. Startling statistics reveal that nearly 15% of patients are readmitted within 30 days of their initial discharge, leading to unfavorable outcomes and increased costs for both patients and hospitals alike.
In a groundbreaking collaboration, NYU Langone Health, the prestigious academic medical center of New York University, has joined forces with NVIDIA experts to develop a cutting-edge solution: a large language model (LLM) that accurately predicts a patient’s risk of 30-day readmission, as well as other critical clinical outcomes.
Introducing NYUTron, the result of this visionary partnership, which has been deployed across the healthcare system’s six inpatient facilities. Today, the remarkable capabilities of NYUTron are showcased in the renowned scientific journal Nature, demonstrating the power of AI-driven insights that enable doctors to identify patients requiring clinical interventions to minimize the likelihood of readmission.
Dr. Eric Oermann, Assistant Professor of Radiology and Neurosurgery at NYU Grossman School of Medicine and a leading collaborator on NYUTron, explains, “When you discharge a patient from the hospital, you don’t expect them to need to return, or you probably should have kept them in the hospital longer. Using analysis from the AI model, we could soon empower clinicians to prevent or rectify situations that put patients at a higher risk of readmission.“
The NYUTron model has already been applied to over 50,000 patients discharged within NYU’s healthcare system, where it seamlessly shares predictions of readmission risk with physicians through email notifications. Building upon this success, Dr. Oermann’s team is now planning a clinical trial to evaluate the effectiveness of interventions based on NYUTron’s analyses in reducing readmission rates.
Addressing the Challenge of Rapid Readmission and Beyond
Monitoring readmission rates within 30 days is a key metric used by the U.S. government to assess the quality of care provided by hospitals. Institutions with high readmission rates face financial penalties, leading to a heightened focus on improving the discharge process.
There are various reasons why recently discharged patients may require readmission, including infections, overprescription of antibiotics, or premature removal of surgical drains. By identifying these risk factors at an earlier stage, doctors can intervene by adjusting treatment plans or providing extended hospital monitoring.
Dr. Oermann emphasizes, “While computational models to predict patient readmission have been available since the 1980s, we treat this as a natural language processing task that necessitates a health system-scale corpus of clinical text. We trained our LLM on the unstructured data of electronic health records to uncover insights that were previously unexplored.“
NYUTron underwent extensive pretraining using a vast dataset comprising more than four billion words of clinical notes from over 400,000 patients within NYU Langone Health. The model’s exceptional accuracy improvement of over 10% surpasses that of a state-of-the-art machine learning model designed for readmission prediction.
Following the successful training of the LLM for the primary use case of 30-day readmission, the team achieved a remarkable feat by developing four additional predictive algorithms in just one week. These encompass predicting the length of a patient’s hospital stay, the likelihood of in-hospital mortality, and the probability of insurance claims being denied.
Optimizing the Journey from Training to Deployment
NYUTron represents an LLM with hundreds of millions of parameters, meticulously trained using the powerful NVIDIA NeMo Megatron framework on a large cluster of NVIDIA A100 Tensor Core GPUs.
Dr. Oermann sheds light on their unique approach, stating, “While current discussions around language models revolve around gargantuan, general-purpose models with billions of parameters trained on complex datasets using numerous GPUs, we opted for medium-sized models trained on highly refined data to fulfill healthcare-specific tasks.”
To ensure efficient inference of the model within real-world hospital environments, the team developed a modified version of NVIDIA Triton, an open-source software, and leveraged the NVIDIA TensorRT software development kit for streamlined AI model deployment.
“When it comes to deploying a model of this nature in a live healthcare environment, efficiency is paramount,” Dr. Oermann affirms. “Triton offers a comprehensive inference framework, delivering the blazing-fast performance required by our model.”
Furthermore, Dr. Oermann’s team discovered that fine-tuning the LLM on-site with data from a specific hospital significantly enhanced its accuracy. This valuable insight enables other healthcare institutions to adopt a similar approach, even if they lack the resources to train a large language model from scratch in-house. By leveraging the cloud’s GPU capabilities, hospitals can readily adopt a pretrained model like NYUTron and fine-tune it using a small sample of their local data.
Conclusion:
The collaboration between NYU Langone Health and NVIDIA signifies a major breakthrough in the healthcare market. The development of NYUTron, a powerful large language model, has the potential to revolutionize patient readmission prediction and improve outcomes. By harnessing the capabilities of AI and natural language processing, hospitals can proactively identify and address factors contributing to readmission, leading to enhanced patient care and reduced costs. This collaboration sets a precedent for future advancements in healthcare-specific AI models, empowering institutions to optimize resource allocation and provide more efficient and personalized care.
