A Thoughtful Revolution: The Integration of Supercomputing and AI at NYU Langone Health

The following is a guest article by Melyssa Fratkin, Communications Chair for SC24.

For decades, healthcare organizations have adopted computing technology to improve operations and support patient care. From desktop computers to tablets, this technology has become essential in modern healthcare, improving the ability to manage patient data, enhance communication, and increase efficiency. But with the rise of Artificial Intelligence (AI), it is clear that supercomputing is poised to become a new backbone of healthcare innovation. To learn about how healthcare organizations are safely and effectively integrating these systems, we spoke with Yin Aphinyanaphongs, MD, PhD, an Assistant Professor in Population Health and the Director of Operational Data Science and Machine Learning at NYU Langone Health, who is helping to lead a “thoughtful revolution” in medicine with generative AI.

“We believe that AI can transform patient care, education, and research,” says Aphinyanaphongs, who is engaged in the development and practical application of AI models across NYU Langone’s entire enterprise, including implementation, integration, and evaluation.

According to Yin Aphinyanaphongs, all AI initiatives at NYU Langone Health are driven and supported centrally by executive and institutional leadership, especially as AI becomes integral to NYU Langone’s mission of advancing clinical care and training future physician-scientists. “Our eventual goal is that all text generated by or for us will be touched by a large language model, whether in our education, clinical, research, or operations missions,” he says. This vision underscores the impact that AI and, consequently, supercomputing is having in healthcare – driving innovation in entirely new ways.

AI is for Everyone

Aphinyanaphongs quickly recognized AI’s potential for revolutionizing healthcare.  “The first time I experienced GPT and other large language models, I realized they would change everything in healthcare,” he recalls. He notes how AI can streamline everything from back-office operations, like legal compliance and finance, by assisting with data management and augmenting some traditional roles. He emphasizes that the institution has taken a democratized approach to development and implementation. “Early on, we solicited ideas from our community, believing that AI is for everyone,” he explains. “We have made a secure, HIPAA-compliant (Health Insurance Portability and Accountability Act) instance of ChatGPT available upon request to our faculty, students, and staff.”

But the back office isn’t the only place where efficiency benefits are being realized, he says. NYU Langone is also using AI to enhance patient care quality, safety, and efficiency. “AI will transform how we care for patients, run hospitals, write research grants, and train medical students,” says Aphinyaphongs, explaining that NYU Langone currently has 31 AI models in various operational areas, including patient care and education.

Meet NYUTron: Clinical Predictions with Domain-Specific AI

AI’s potential in medicine is vividly demonstrated by NYUTron (pronounced neutron), a generative large language model developed with leadership from Eric Oermann, MD, Associate Professor in the Departments of Neurosurgery and Radiology at NYU Grossman School of Medicine. NYUTron is a BERT (Bidirectional Encoder Representations from Transformers)-style model, not an auto-regressive model like a GPT. This means its primary function is to understand words in a bidirectional manner, making it particularly useful for tasks where clearly understanding the full context is important. “You can’t chat with it, but it can handle classifications and fine-tuning,” Aphinyaphongs explains.

NYUTron was trained on 7.25 million patient-care notes and 4.1 billion words, incorporating data spanning nearly a decade. Built in roughly three weeks, the 109-million-parameter model is being put to the test with patients. Aphinyanaphongs explains that “our unique contribution was evaluating its effectiveness for real clinical tasks, such as predicting inpatient hospital mortality, computing the comorbidity index, 30-day all-cause readmission, length of stay, and payer denial at discharge. We showed state-of-the-art performance for these tasks and are now building clinical trials around these models.” In the inpatient mortality use case, the goal is to use NYUTron to improve patient care by providing early warnings and facilitating important conversations about patient care preferences and, when appropriate, end-of-life planning.

Assessing Accountability in AI

But the implementation isn’t all smooth sailing, admits Aphinyanaphongs. Some tasks, like predicting 30-day readmissions, are very challenging. He shares that creating effective workflows is a big part of the challenge. “The model’s performance is only part of the equation,” he says. Aphinyanaphongs highlighted that workflow integration and broader implementation of AI in healthcare pose key challenges with no easy answers: “Effective communication and convincing healthcare providers of the model’s accuracy are crucial.” Another challenge, he says, is not just creating accurate models but also ensuring they can be effectively incorporated into daily medical practices: “It’s one thing to have a model that predicts something, but another thing entirely to determine if that prediction is strong enough to prompt action from a provider. How do you integrate these predictions into the workflow? It’s something that takes time and effort to build confidence.”

A big part of it comes down to building a culture that understands what AI is and is not useful for and building accountability into the process.

“First, everyone needs to be aware that these models can hallucinate, and this needs to be emphasized continuously,” he explains. “Understanding why a model made a specific prediction is crucial.”

Emphasizing the importance of model transparency, he stresses NYU Langone’s perspective that AI should be used to augment human decision-making, leveraging the computational power to enhance patient care while relying on human judgment to interpret and act on AI predictions. “Any AI recommendation is just a data point, like a lab result, meant to augment human capabilities not replace them,” Aphinyanaphongs says.

Vital System Support

Foundational to NYU Langone’s AI endeavors is its Ultraviolet HPC cluster, a supercomputing system that supports a range of computationally intensive research programs, including (but not limited to) training many of their sophisticated AI models.

With 197 nodes, 646 GPUs, 13,736 CPUs, 15 petabytes of active storage, and currently exploring “cloud bursting” to commercial compute clouds, Ultraviolet is a top-tier computational system, currently ranked 397 in the Top500, and is ranked 16th among U.S. academic systems*. “It is the only system in the Top500 dedicated to healthcare and the life sciences, demonstrating that we are still in the early stages of the use of supercomputing in healthcare,” John Speakman, Assistant Vice President, Research IT at NYU Langone Health, says.

Beyond NYUTron, the Ultraviolet HPC cluster is the backbone for the development of more than 30 live AI models, enhancing backend and frontend workflows and improving patient care outcomes. For instance, AI models ensure that critical medications are ordered and documented correctly, prioritize urgent patient messages, and draft responses to medical advice and medication refill requests. These applications demonstrate the broad impact of AI across various aspects of healthcare delivery. The cluster also manages and processes massive datasets, including millions of patient records, vital sign measurements, documentation notes, and test results. By supporting these diverse functions, the Ultraviolet HPC cluster not only enhances the capabilities of NYUTron but also powers various AI and data-intensive initiatives central to advancing healthcare at NYU Langone Health.

Combining Technology and Expertise: The Path Forward

While the Ultraviolet HPC cluster provides considerable computational muscle, combining this technology with the expertise and innovative spirit at NYU Langone is what truly drives progress. Thoughtful integration of AI into clinical practices, strongly supported by organizational leadership and physician-scientists ensures these powerful tools are used effectively to improve patient care.

“The one word I can leave you with that encompasses our approach is ‘thoughtful,’” Aphinyanaphongs says. “There are so many unknowns in this space, and while there’s a push for speed and quickness, we aim to be incredibly thoughtful with our research.”

Connect, Learn, and Innovate at SC24

The path to implementing and maximizing advancements in AI and supercomputing, like at NYU Langone Health, is ultimately navigated by making connections and learning from others. SC24 offers an unparalleled opportunity to expand your knowledge and enrich your experiences within the HPC community. Attendees can engage with technical presentations, papers, workshops, tutorials, posters, and Birds of a Feather sessions – all designed to showcase the latest innovations and practical applications in AI and HPC. The conference offers a unique platform, where experts from leading manufacturers, research organizations, industry, and academia come together to share insights and advancements that are driving the future. Join the community for a week of innovation at SC24 in Atlanta, November 17-22, 2024.

About Melyssa Fratkin

Melyssa Fratkin is serving as the Communications Chair for the SC24 conference. In her day job, she is the Industrial Programs Director at the Texas Advanced Computing Center (TACC) at The University of Texas at Austin. In this role, she oversees TACC’s corporate partnerships and government relations activities. Melyssa focuses on developing and managing strong collaborations between industry and academia with the Science & Technology Affiliates for Research (STAR) program.

Melyssa is also the founding co-chair of Texas Women in HPC, an organization aimed at raising awareness and broadening diversity in HPC by supporting women and under-represented populations in high performance computing in industry, academia, and government across Texas. She also serves as Vice President of the AMD HPC User Group, and Communications Committee Chair for the Coalition for Academic Scientific Computation (CASC). Melyssa received a BA from Rutgers University and MBA from the Robert H. Smith School of Business at the University of Maryland.

*Ranking at the time of release.

The following is a guest article by Melyssa Fratkin, Communications Chair for SC24.

For decades, healthcare organizations have adopted computing technology to improve operations and support patient care. From desktop computers to tablets, this technology has become essential in modern healthcare, improving the ability to manage patient data, enhance communication, and increase efficiency. But with the rise of Artificial Intelligence (AI), it is clear that supercomputing is poised to become a new backbone of healthcare innovation. To learn about how healthcare organizations are safely and effectively integrating these systems, we spoke with Yin Aphinyanaphongs, MD, PhD, an Assistant Professor in Population Health and the Director of Operational Data Science and Machine Learning at NYU Langone Health, who is helping to lead a “thoughtful revolution” in medicine with generative AI.

“We believe that AI can transform patient care, education, and research,” says Aphinyanaphongs, who is engaged in the development and practical application of AI models across NYU Langone’s entire enterprise, including implementation, integration, and evaluation.

According to Yin Aphinyanaphongs, all AI initiatives at NYU Langone Health are driven and supported centrally by executive and institutional leadership, especially as AI becomes integral to NYU Langone’s mission of advancing clinical care and training future physician-scientists. “Our eventual goal is that all text generated by or for us will be touched by a large language model, whether in our education, clinical, research, or operations missions,” he says. This vision underscores the impact that AI and, consequently, supercomputing is having in healthcare – driving innovation in entirely new ways.

AI is for Everyone

Aphinyanaphongs quickly recognized AI’s potential for revolutionizing healthcare.  “The first time I experienced GPT and other large language models, I realized they would change everything in healthcare,” he recalls. He notes how AI can streamline everything from back-office operations, like legal compliance and finance, by assisting with data management and augmenting some traditional roles. He emphasizes that the institution has taken a democratized approach to development and implementation. “Early on, we solicited ideas from our community, believing that AI is for everyone,” he explains. “We have made a secure, HIPAA-compliant (Health Insurance Portability and Accountability Act) instance of ChatGPT available upon request to our faculty, students, and staff.”

But the back office isn’t the only place where efficiency benefits are being realized, he says. NYU Langone is also using AI to enhance patient care quality, safety, and efficiency. “AI will transform how we care for patients, run hospitals, write research grants, and train medical students,” says Aphinyaphongs, explaining that NYU Langone currently has 31 AI models in various operational areas, including patient care and education.

Meet NYUTron: Clinical Predictions with Domain-Specific AI

AI’s potential in medicine is vividly demonstrated by NYUTron (pronounced neutron), a generative large language model developed with leadership from Eric Oermann, MD, Associate Professor in the Departments of Neurosurgery and Radiology at NYU Grossman School of Medicine. NYUTron is a BERT (Bidirectional Encoder Representations from Transformers)-style model, not an auto-regressive model like a GPT. This means its primary function is to understand words in a bidirectional manner, making it particularly useful for tasks where clearly understanding the full context is important. “You can’t chat with it, but it can handle classifications and fine-tuning,” Aphinyaphongs explains.

NYUTron was trained on 7.25 million patient-care notes and 4.1 billion words, incorporating data spanning nearly a decade. Built in roughly three weeks, the 109-million-parameter model is being put to the test with patients. Aphinyanaphongs explains that “our unique contribution was evaluating its effectiveness for real clinical tasks, such as predicting inpatient hospital mortality, computing the comorbidity index, 30-day all-cause readmission, length of stay, and payer denial at discharge. We showed state-of-the-art performance for these tasks and are now building clinical trials around these models.” In the inpatient mortality use case, the goal is to use NYUTron to improve patient care by providing early warnings and facilitating important conversations about patient care preferences and, when appropriate, end-of-life planning.

Assessing Accountability in AI

But the implementation isn’t all smooth sailing, admits Aphinyanaphongs. Some tasks, like predicting 30-day readmissions, are very challenging. He shares that creating effective workflows is a big part of the challenge. “The model’s performance is only part of the equation,” he says. Aphinyanaphongs highlighted that workflow integration and broader implementation of AI in healthcare pose key challenges with no easy answers: “Effective communication and convincing healthcare providers of the model’s accuracy are crucial.” Another challenge, he says, is not just creating accurate models but also ensuring they can be effectively incorporated into daily medical practices: “It’s one thing to have a model that predicts something, but another thing entirely to determine if that prediction is strong enough to prompt action from a provider. How do you integrate these predictions into the workflow? It’s something that takes time and effort to build confidence.”

A big part of it comes down to building a culture that understands what AI is and is not useful for and building accountability into the process.

“First, everyone needs to be aware that these models can hallucinate, and this needs to be emphasized continuously,” he explains. “Understanding why a model made a specific prediction is crucial.”

Emphasizing the importance of model transparency, he stresses NYU Langone’s perspective that AI should be used to augment human decision-making, leveraging the computational power to enhance patient care while relying on human judgment to interpret and act on AI predictions. “Any AI recommendation is just a data point, like a lab result, meant to augment human capabilities not replace them,” Aphinyanaphongs says.

Vital System Support

Foundational to NYU Langone’s AI endeavors is its Ultraviolet HPC cluster, a supercomputing system that supports a range of computationally intensive research programs, including (but not limited to) training many of their sophisticated AI models.

With 197 nodes, 646 GPUs, 13,736 CPUs, 15 petabytes of active storage, and currently exploring “cloud bursting” to commercial compute clouds, Ultraviolet is a top-tier computational system, currently ranked 397 in the Top500, and is ranked 16th among U.S. academic systems*. “It is the only system in the Top500 dedicated to healthcare and the life sciences, demonstrating that we are still in the early stages of the use of supercomputing in healthcare,” John Speakman, Assistant Vice President, Research IT at NYU Langone Health, says.

Beyond NYUTron, the Ultraviolet HPC cluster is the backbone for the development of more than 30 live AI models, enhancing backend and frontend workflows and improving patient care outcomes. For instance, AI models ensure that critical medications are ordered and documented correctly, prioritize urgent patient messages, and draft responses to medical advice and medication refill requests. These applications demonstrate the broad impact of AI across various aspects of healthcare delivery. The cluster also manages and processes massive datasets, including millions of patient records, vital sign measurements, documentation notes, and test results. By supporting these diverse functions, the Ultraviolet HPC cluster not only enhances the capabilities of NYUTron but also powers various AI and data-intensive initiatives central to advancing healthcare at NYU Langone Health.

Combining Technology and Expertise: The Path Forward

While the Ultraviolet HPC cluster provides considerable computational muscle, combining this technology with the expertise and innovative spirit at NYU Langone is what truly drives progress. Thoughtful integration of AI into clinical practices, strongly supported by organizational leadership and physician-scientists ensures these powerful tools are used effectively to improve patient care.

“The one word I can leave you with that encompasses our approach is ‘thoughtful,’” Aphinyanaphongs says. “There are so many unknowns in this space, and while there’s a push for speed and quickness, we aim to be incredibly thoughtful with our research.”

Connect, Learn, and Innovate at SC24

The path to implementing and maximizing advancements in AI and supercomputing, like at NYU Langone Health, is ultimately navigated by making connections and learning from others. SC24 offers an unparalleled opportunity to expand your knowledge and enrich your experiences within the HPC community. Attendees can engage with technical presentations, papers, workshops, tutorials, posters, and Birds of a Feather sessions – all designed to showcase the latest innovations and practical applications in AI and HPC. The conference offers a unique platform, where experts from leading manufacturers, research organizations, industry, and academia come together to share insights and advancements that are driving the future. Join the community for a week of innovation at SC24 in Atlanta, November 17-22, 2024.

About Melyssa Fratkin

Melyssa Fratkin is serving as the Communications Chair for the SC24 conference. In her day job, she is the Industrial Programs Director at the Texas Advanced Computing Center (TACC) at The University of Texas at Austin. In this role, she oversees TACC’s corporate partnerships and government relations activities. Melyssa focuses on developing and managing strong collaborations between industry and academia with the Science & Technology Affiliates for Research (STAR) program.

Melyssa is also the founding co-chair of Texas Women in HPC, an organization aimed at raising awareness and broadening diversity in HPC by supporting women and under-represented populations in high performance computing in industry, academia, and government across Texas. She also serves as Vice President of the AMD HPC User Group, and Communications Committee Chair for the Coalition for Academic Scientific Computation (CASC). Melyssa received a BA from Rutgers University and MBA from the Robert H. Smith School of Business at the University of Maryland.

*Ranking at the time of release.