A very large healthcare AI model likened to GPT for health has secured data and processing needed for its next phase of development.
Nightingale AI, led by Imperial College London, will draw on large UK, EU and US health datasets to help scientists and medical specialists with applications such as developing treatments, diagnosing patients and monitoring illness.
The project has received a boost after the UK government's Sovereign AI Unit awarded it one million hours of GPU compute time on the world's eleventh fastest supercomputer, Isambard-AI in Bristol.
Meanwhile, Imperial and California-based Nightingale AI partner the Children's Hospital of Orange County (CHOC), part of Rady Children's Health, have announced a plan to boost AI for children's diseases that have been typically under-served by AI by working on anonymised electronic patient records.
The growing integration of data and compute time at scale will help the Nightingale AI team pursue work to train a 'world' model that could help researchers and clinicians answer a wide range of health-related questions by finding patterns in large datasets.
Reasoning using specialised data
Unlike models that use written language as the basis for reasoning, Nightingale AI will learn from multi-modal healthcare data. This means it will be designed to read data as varied as medical records, lab results, X-rays, genetic and imaging data and published medical research. It is therefore expected to be far better able to reason about patients, providing medical insights that are novel and tailored to healthcare and biomedicine.
Nightingale AI will learn from anonymised medical data without sharing any of the data that was used to create it.
Compute and data
Isambard-AI is a recently launched £225 million supercomputer funded by the UK government and hosted at the University of Bristol that will provide Nightingale AI with initial compute time equivalent to the processing that OpenAI used to train GPT3, the first commercially deployed large language model.
Professor Aldo Faisal, Director of the UKRI Centre for Doctoral Training in AI for Healthcare (AI4Health), will lead a team of doctoral and PhD researchers. They will train Nightingale AI using data provided partly by CHOC, a leading US paediatric healthcare system that has previously worked with Professor Faisal on AI Clinician, an AI tool it has used to help paediatric doctors make complex decisions when treating critically ill children.
A new wave of medical advances
Based in Imperial's School of Convergence Science in Human and Artificial Intelligence, Nightingale AI is led by Imperial in partnership with leading universities in the UK and Europe. It is receiving funding by the UKRI AI Hub in Generative Models, the Horizon Europe programme DVPS, and the AI4Health.
As a non-profit initiative, Nightingale AI will seek to support the development of useful medicines and health technologies while remaining independent of commercial interests.
The initiative was recognised by Nvidia during the recent state visit to the UK by the President of the United States of America, when the GPU company named it among UK AI innovations to use the company's chips.
Professor Faisal said: "By securing access to large scale data and the full compute power of Isambard-AI, we're ready to unlock a completely new quality of AI for healthcare, driving medical advances but also being able to offer every individual patient a tailored medical advice, treatment and referrals.:
Nightingale AI co-lead Dr Marek Rei in Imperial's Department of Computing, an expert in large language models, said : "While there are many language models that will try to answer some medical questions, we are building a model designed specifically for combining knowledge from different types of medical data."
