Medical training’s AI leap: How agentic RAG, open-weight LLMs and real-time case insights are shaping a new generation of doctors at NYU Langone

Patient data records can be curled and sometimes incomplete, meaning that doctors don’t always have all the information they need to provide. Beyond that, medical professionals are unable to keep up with case studies, research papers, trials and other barriers to cutting-edge development.

Headquartered in New York City NYU Langone Health A new approach has been proposed to address these challenges of the next generation of doctors.

Including the Academic Medicine Center at NYU Grossman School of Medicine and NYU Long Island School of Medicine, as well as six inpatient hospitals and 375 outpatients – the Large Language Model (LLM) consultant developed.

Each night, the model processes electronic health records (EHRs), matches them with relevant research, diagnostic techniques, and basic background information, and then provides residents with concise, tailored emails the next morning. This is the fundamental component of NYU Langone’s groundbreaking approach to medical education – what it says is the use of “precision medical education” AI and data Provides highly customized student journeys.

“The concept of precision needed in health care” Marc Triola, associate dean of educational informatics and director of the Langone Health Institute for Medical Education at NYU, told VentureBeat. “Obviously, there is evidence that AI can overcome the health care system. There are many cognitive biases, errors, waste and inefficient cognitive biases that can improve diagnostic decisions. ”

How NYU Langone uses Llama to enhance patient care

NYU Langone is using the latest version of Llama-3.1-8B-Instruct and open source open models Color Vector database Search for a generation (rag). But it’s more than just accessing documents – the model goes beyond rags and actively adopts search and other tools to discover the latest research documents.

Each night, the model is connected to the facility’s EHR database and extracts medical data for patients seen in Langone the day before. It then searches for basic background information about diagnosis and medical conditions. Using the Python API, the model also searches related medical literature PubMedThere are “millions of papers” among them, Triola explained. LLM screened through reviews, in-depth reference papers and clinical trials, selected several seemingly most relevant content and “fused them all together.”

Early the next morning, medical students and residents of internal medicine, neurosurgery and radiation oncology received a personalized email with detailed patient summary. For example, if a patient with congestive heart failure the day before was examined, the email will provide further study on the basic pathophysiology of heart disease and information about the latest therapies. It also provides self-study questions and AI planning Medical Literature. Additionally, it may provide instructions for steps residents may take or actions or details that may be ignored.

“We have received great feedback from students, residents and teachers on how to keep them up to date with friction and how they fit in the way they choose their patients’ care plans,” Triola said.

For him personally, a key indicator of success is that when the system interruption stopped emails for several days, teachers and students complained that they did not receive the morning excavation they depended on.

“Because we sent these emails before the doctor started the round – one of the craziest and busiest times of their day – make them notice they didn’t receive these emails and miss them thinking as part of them Great,” he said.

Change the industry with precise medical education

This complex AI retrieval system is the basis of Nyu Langone’s Precision medical education model, which Triola explains is based on “higher density, frictionless” digital data, AI and strong algorithms.

Over the past decade, the institution has collected a lot of data about students – their performance, the environment they are caring for, and the EHR notes they write, Clinical decision-making The way they are doing and the way they reason through patient interaction and care. Additionally, NYU Langone has a large catalogue in all resources for medical students, whether it’s videos, self-study or exam questions, or online learning modules.

The success of the project is also attributed to the simplified architecture of the healthcare organization: it has a centralized, a single data warehouse for healthcare and a single data warehouse for education, allowing Langone to marry its various data resources.

Chief Medical Information Officer Paul Testa noted that a great AI/ML system cannot be implemented without big data, but “if you sit in a silo in the system, it’s not the easiest thing to do.” Healthcare systems can be big, But it serves as “one patient, one record, one standard”.

AI generation allows Nyu Langone to get out of “a certain level of all-round” education

As Triola said, the main question his team has been looking for is: “How do they connect the diagnosis, the individual student’s background and all of this learning material?”

“Suddenly, we had a great key to unlock this: the generated AI,” he said.

This enables schools to get out of the “crystal” model that has always been the norm, and a unique approach is needed for students whether they intend to become, for example, neurosurgeons or psychiatrists.

It is important that students receive a tailored education throughout the education process, as well as a “light education” that adapts to their needs, he said. But you can’t just tell teachers to “spend more time with each student” – it’s impossible for humans.

“Our students are hungry for this because they recognize that this is a period of rapid change in medicine and generating AI,” Triola said. “It will definitely change … what it means to be a doctor.”

As a role model for other medical institutions

It’s not that there are no challenges along the way. It is worth noting that the technical team has been working through the model “immature”.

As Triola points out: “It’s fascinating, how inflated and accurate their embedded knowledge is, sometimes limited. It will be perfect 99 consecutive times, predictable, and then the 100th will make interesting choices.”

For example, in the early stages of development, LLM cannot distinguish between ulcers on the skin and ulcers in the stomach, which is “conceptually irrelevant”. Since then, his team has focused on rapid refining and rooting, and the result is “excellent.”

In fact, his team is so confident about the stack and the process that they think it can be a good example for others to follow. “We love open source and open weight because we want to get to the point where, ‘Hey, other medical schools, many of them don’t have a lot of resources and you can do it at a cheap price,’ Triola explained.

Testa agrees: “Is it reproducible? Is this something we want to spread? Absolutely, we want to spread it in health care.”

Reevaluating the practice of “Sacred Uneasy” in medicine

It is understandably important to note that throughout the industry there is a lot of attention to the subtle biases that may be baked into AI systems. However, Triola points out that this is not a huge problem in this use case, as it is a relatively simple task for AI. “It’s searching, it’s selected from the list, it’s summarizing,” he said.

Instead, one of the biggest surface problems is around barrier-free or cabinets. Here is a correlation: a certain year correlation may remember learning cursive in elementary school – but they may have forgotten the skill because they discovered a rare opportunity as adults. Now, it is almost outdated and rarely taught in today’s primary education.

Triola notes that being a “sacred and inviolable” part of a doctor, some people can give it to an AI or digital system “in any way, shape or form.” For example, it is believed that as long as young doctors are not in a clinical setting, young doctors should actively study and nose in the latest literature. However, Triola emphasizes the amount of medical knowledge available today and the “fanatic speed” of clinical medicine, requiring a different way of doing things.

In terms of researching and retrieving information, he noted: “AI does better, which is an unsettling fact that many people are hesitating.”

Instead, he hypothesized: “Suppose this will give doctors a superpower and find out the co-pilot relationship between human and artificial intelligence, rather than the competitive relationship of who will do what.”