Can an AI large language model transform health care?

Med42 seeks to revolutionise clinical experience for patients and doctors

Med42, a new clinical large language model (LLM), provides answers to short and long-form medical questions. Photo: M42

Since October, Abu Dhabi’s Department of Health has been assisting with the eventual implementation of a new clinical generative AI large language model (LLM), Med42, across the emirate’s healthcare ecosystem.

“It’s designed to serve as an assistant to patients, physicians, nursing staff and other healthcare professionals,” said Shadab Khan, head of AI and applied sciences at M42, Abu Dhabi’s technology-enabled healthcare company which created Med42.

“We think Med42 will be a great companion to patients,” he added, while also noting that the model was in a test phase.

“We’re planning to assess the model’s safety, reliability and bias in its outputs,” he said.

“Once we're satisfied that it's safe for application in this highly regulated field of health care, we'll be able to offer it to patients and other individuals in our network.”

M42 said the 70 billion parameter LLM was trained on the company's industry-leading curated data set of medical knowledge.

In a demo video provided to The National, a hypothetical doctor asked Med42 a question: “I’ve been managing an increasing number of patients with coexisting heart failure and renal dysfunction … can you elucidate the mechanisms underpinning the bidirectional relationship between renal dysfunction and heart failure and how they might influence our therapeutic approach?”

“The bidirectional relationship between renal dysfunction and heart failure is complex and multifactorial,” Med42 answers.

“On one hand, heart failure can lead to renal dysfunction due to reduced cardiac output, systemic hypotension, and increased sympathetic activation which can cause renal vasoconstriction and decreased glomerular filtration rate,” its response reads in part.

To help train the LLM, one of the world’s most powerful supercomputers, Condor Galaxy 1, was used.

“It was quite an integral piece in ensuring we were able to hit our milestones on time,” said Mr Khan, referring to the assistance of the supercomputer.

A team from the Mohamed bin Zayed University of Artificial Intelligence, the world's first dedicated artificial intelligence university, also evaluated the Med42 LLM for accuracy.

According to M42, the Med42 LLM achieved a 72 per cent score on the US medical licensing examination sample questions, outperforming other language models.

“It was a really proud moment for us,” said Mr Khan.

“But at the same time, we were also grounded in the fact that achieving a score on an exam is one thing, and delivering actual value to our end users is quite another.”

The model currently is designed and tested to understand English, although Mr Khan said it does understand other languages to a degree.

“Arabic capability is an important feature on a road map that’s coming shortly,” he said.

The LLM has also been made available for download on Hugging Face, an online machine learning community which collaborates on models, data sets and applications, to allow for widespread testing and scientific assessment.

M42 is not alone in efforts to utilise artificial intelligence in the medical field. Google recently unveiled MedLM, which the company describes as a “foundation of models fine-tuned for healthcare industry use cases”, available to Google Cloud customers in the US.

MedLM is loosely based on Google’s initial research into medical LLMs, such as Med-PaLM 2, which it has been testing with various healthcare organisations.

LLMs have been instrumental in the growth and utilisation of AI, helping to produce content by training on large amounts of data, which in turn, enables users to get answers and automate tasks in a fraction of the time previously needed.