Artificial Intelligence has become one of the favourite topics and tools like DALL E 2, ChatGPT and more are helping the users across the globe in a variety of tasks. According to a study, a new AI tool that can generate doctors’ notes so perfectly that two physicians couldn’t tell the difference may soon open the door for AI to help health care personnel with breakthrough efficiencies.
In this proof-of-concept study, doctors examined patient notes, some of which were authored by real medical professionals and others by the new AI system. Only 49% of the time, the doctors were able to identify the correct author. A group of 19 researchers from the University of Florida and NVIDIA trained supercomputers to create medical records using a new model called GatorTronGPT, which works in a similar way as ChatGPT.
READ: Google’s New AI Tool Will Now Help Brands, Creative Agencies To Make More Impactful Ads; Here’s How
Hugging Face is an open-source AI website where more than 430,000 people have downloaded the free versions of GatorTron models. Lead author Yonghui Wu of the University of Florida’s department of health outcomes and biomedical informatics claims that GatorTron models are the only models on the site that are suitable for clinical research.
“Everyone in the medical field is discussing these models. Several facets of medical research and healthcare can be powered by the innovative AI models GatorTron and GatorTronGPT. However, to develop them, enormous amounts of data and processing power are needed. We are appreciative of NVIDIA’s loan of this supercomputer, HiPerGator, to investigate the possibilities of AI in healthcare,” Wu said.
The researchers created a huge language model for this study, which was published in the journal npj Digital Medicine. This model enables computers to emulate natural human discourse. Standard writing and conversations are a good fit for these models; nevertheless, medical records provide extra challenges due to their high technicality and the requirement to maintain patient privacy. Digital health records are not searchable on Google or accessible on Wikipedia.
The researchers kept 82 billion valuable medical words while utilising the medical records of two million people to get over these challenges. They trained the GatorTronGPT model to analyse the medical data using GPT-3 architecture, or Generative Pre-trained Transformer, a type of neural network architecture, by combining this collection with another dataset of 195 billion words. GatorTronGPT was able to produce clinical writing that resembles doctor’s notes as a result.
One of the many potential applications for a medical GPT is to replace the tediousness of paperwork with notes that are taken by AI and then transcribed. Programmers spend weeks equipping supercomputers with clinical vocabulary and language usage based on billions upon billions of words so that an AI tool may write on a par with human writing.
