New ChatGPT-like AI tool’s remarkable ability to mimic doctors' notes stuns researchers

A recent study revealed that a novel artificial intelligence (AI) computer program can craft doctors' notes so convincingly that distinguishing them from those written by human physicians proved challenging.  In this proof-of-concept investigation, physicians evaluated patient notes, some composed by actual medical professionals and others generated by the new AI programme. Astonishingly, the physicians accurately identified the true author only 49 per cent of the time.

A collaborative team of 19 researchers from NVIDIA and the University of Florida dedicated efforts to train supercomputers in producing medical records based on an innovative model named GatorTronGPT, which operates similarly to ChatGPT.

The freely accessible GatorTron models have garnered over 430,000 downloads via Hugging Face, an open-source AI platform. Notably, GatorTron models stand as the site's exclusive offerings available for clinical research, as noted by lead author Yonghui Wu from the University of Florida's Department of Health Outcomes and Biomedical Informatics.

Wu expressed, "Within the healthcare realm, these models have sparked significant conversations. GatorTron and GatorTronGPT represent distinctive AI models poised to impact numerous facets of medical research and healthcare. However, their construction demands substantial data and extensive computational resources. We are fortunate to leverage the capabilities of the NVIDIA supercomputer, HiPerGator, in exploring AI's potential within healthcare."

The research, published in the journal npj Digital Medicine, involved crafting a comprehensive language model enabling computers to emulate natural human language.

While these models excel in conventional writing or conversations, adapting them to handle medical records introduces additional challenges, encompassing the imperative need to safeguard patients' privacy and the requirement for highly technical precision. Medical records within the digital domain cannot be sourced via search engines like Google or from platforms such as Wikipedia.

To surmount these hurdles, researchers utilised health-related medical records from a cohort of two million patients, encompassing an array of 82 billion pertinent medical terms. Integrating this dataset with an additional collection of 195 billion words, they trained the GatorTronGPT model to analyse medical data using GPT-3 architecture, a variant of neural network architecture. This facilitated GatorTronGPT's ability to compose clinical text resembling doctors' notes.

Among the various potential applications for a medical GPT, one promising notion involves replacing the arduous task of documentation with notes transcribed and recorded by AI.

Achieving such alignment between AI-generated text and human writing necessitates extensive programming efforts, with programmers dedicating weeks to equip supercomputers with clinical vocabulary and language patterns derived from billions upon billions of words.

Also Read Generative AI to replace 46% of time spent on accounting tasks in India; impact of AI on white vs blue-collar jobs