AI Leap of Medical Training: How the Agency rag, Open-Weight LLMS and Insights in real time form a new generation of doctors at NYU Langone

Patient data records can be confused and sometimes incomplete, which means that doctors do not always have all the information they need easily. To this is the fact that medical professionals cannot be up to date with the partition of cases, research documents, tests and other avant -garde developments coming out of the industry.

New York -based NYU Langone Health He came up with a new approach to address these challenges for the next generation of doctors.

The Academic Medical Center – which consists of the NYU Grossman Medical School and the NYU Grossman Long Island Medical School, as well as six hospital hospitals and 375 outpatients – developed a large language model (LLM), which serves as a respected research satellite S

Each night, the model processes e -health records (EHR), comparing them with appropriate research, diagnostic tips and basic information, which then provides in the concise, adapted emails to residents the next morning. This is an elementary part of the Nyu Langone Pioneer approach to the Medical School – what he calls a “precision medical education” he uses You have the date To provide high personalized trips to students.

“This concept of” precision in everything “is needed in healthcare,” said Mark Triola, Assistant Professor of Educational Informatics and Director of the Institute for Innovation in Medical Education at NYU Langone Health, “Venturebeat told. “Obviously, there is evidence that AI can overcome many of the cognitive biases, mistakes, waste and inefficiency in the healthcare system that it can improve diagnostic decision -making.”

How Nyu Langone uses Llama to improve patient care

Nyu Langone uses an open-weight model built on the latest version of Llama-3.1-8B-Instruct and open code Chromium vector database generation extracted generation (Rag). But this is not just access to documents – the model exceeds the rag, actively uses search and other tools to find the latest research documents.

Each night, the model connects to the EHR database of the facility and removes medical data on patients observed in Langone the previous day. He then seeks basic information on diagnoses and medical conditions. Using Python API, the model also searching for related medical literature in PUBMEDWho has “millions and millions of documents,” Triola explained. LLM reviews reviews, deep diving documents and clinical trials, choosing a few at first glance the most appropriate and “puts it all in a nice email.”

At the beginning of the next morning, students in medicine and internal medicine, neurosurgery and radiation oncology residents receive a personalized email with detailed patient summaries. For example, if a patient with congestive heart failure has been inspected the previous day, the email will provide refreshing the basic pathophysiology of heart conditions and information about the most treatments. It also offers questions about self -study and AI medical literatureS In addition, this can give the director for the steps that residents could take further or actions or details that they may have neglected.

“We have received great feedback from students, from residents and from the faculty about how this without friction keeps them up -to -date, how they include this in the way they make a choice for the patient’s care plan,” the trio said.

A key indicator of success for him was personally when the system interruption stopped the emails for a few days – and the teachers and students complained that they did not receive the morning pressures they had come to rely.

“So thinking was great,” he said.

Conversion of industry with precision medical education

This sophisticated AI extracting system is a major for the NYU Langone medical education model, which Triola explained, is based on “higher density, without rubbing” digital data, AI and strong algorithms.

The institution has collected huge amounts of data in the last decade for students – their presentation, the environment in which they take care of patients, EHR notes that they write, Clinical solutions They also do the way they think through interactions and care for the patient. In addition, NYU Langone has a huge catalog of all resources available to medicine students, whether these are videos, self -study questions or exams, or online learning modules.

The success of the project is also thanks to the optimized architecture of the medical establishment: it boasts a centralized, one healthcare warehouse and one education warehouse that allows Langone to marry its various data resources S

The chief medical officer Paul Test noted that great AI/ML systems are not possible without great data, but “is not the easiest thing if you are sitting on unwanted data in silos throughout your system.” The medical system may be large, but it works as “one patient, one record, one standard”.

Gen Ai that allows Nyu Langone to move away from the one -size -all education for all

As the Triola says, the main question that his team is looking for is: “How do they associate the diagnosis, the context of the individual student and all these teaching materials?”

“Suddenly we have this great key to unlock this: generative AI,” he said.

This enabled the school to move away from the one-size model for everyone, which is the norm, whether the students intended to become, for example, a neurosurgeon or a psychiatrist-significant disciplines that require unique approaches.

It is important for students to receive consistent education throughout their school, as well as “educational pressures” that adapt to their needs, he said. But you can’t just tell teachers to “spend more time with each individual student” – it’s impossible human.

“Our students are hungry for this because they admit that this is a period of high speed of change in medicine and generative AI,” Triola said. “It will absolutely change … What it means to be a doctor.”

Serves as a model for other medical institutions

Not that there were no challenges along the way. In particular, the technical teams work through the immaturity model.

As Triola noted: “It is fascinating how expansive and accurate their built -in knowledge is, and sometimes how limited. It will work perfectly, predictably, 99 times in a row, and then in the 100th time it will make an interesting set of choices. “

For example, at the beginning of development, LLMS cannot distinguish between skin ulcers and ulcer in the stomach, which “are not at all related to conceptual attitude,” Triola explained. Since then, his team has focused on rapid refining and grounding and the result is “remarkable”.

In fact, his team is so confident in the stack and the process that they believe it can serve as a great example of following others. “We preferred an open code and an open weight because we wanted to get to the point that we could say,” Hey, other medical schools, many of which do not have many resources, you can do this to the cheap, “Triola explained.

The test agreed: “Is it reproducible? Is this something we want to spread? We absolutely want to spread it to healthcare. “

Sacrosanct reassessment practices in medicine

Understandably, there is a great deal of concern in the Indi about nuanced bias that can be baked in AI systems. However, Triola said that this is not a huge concern in this case, as this is a relatively clear task for AI. “Looking for, chooses from a list, summarizes,” he noted.

More recently, one of the biggest problems with the occurrence is about inability or attachment. Here’s a correlation: those of a particular harvest can remember learning a ittel in elementary school – but they have probably forgotten the skills because they have found a rare reason to use it in their life as adults. He is now almost outdated, rarely taught in today’s elementary education.

Triola pointed out that there are “sacred” parts of being a doctor and some are resistant to refer to those of AI or digital systems “in any way, shape or shape”. For example, there is a perception that young doctors should actively examine and descend into the latest literature when they are not in a clinical setting. But the amount of medical knowledge available today and the “frenetic pace” of clinical medicine requires a different way of doing things, the trio stressed.

As for the study and retrieval of information, he noted: “AI makes it better and this is an uncomfortable truth in which many people hesitate to believe.”

Instead, he stated: “Let’s say that this will give the superpowers to the doctors and we will understand the connection of the pilot between man and AI, not the competitive relationships of who will do what.”