March 22, 2023 – Flare Capital Partners Expert Roundtable on Generative AI in Healthcare.
Conversation flowed between the panel of industry leaders, including Atropos CEO Brigham Hyde, Ph.D. As he put it:
”All of us need to recognize that user experience has fundamentally changed in search. When you look at the adoption of ChatGPT… the standard for user experience is fundamentally different. That’s something we believe at Atropos, and when we work with physicians, we provide a simple chat interface. You should be able to just ask a question like you’re talking to a colleague. As well as receive a summarized and contextualized output that’s easy to understand. Much of that has happened, we need to act as if it’s the new truth. That has huge implications for chatbots, SEO, marketing and advertising, and across the board. That’s a good thing in healthcare, where UX for physicians hasn’t always been great. At Atropos we sit on hundreds of millions of patient records and provide the best insight on the best datasets.”
He continued, “My biggest challenge with ChatGPT to date is that LLMs are largely trained on published literature in the health sciences. The clinical trials, articles, and maybe popular journalism or editorials. And they’re getting to be great at it. But we know right now that we don’t have enough evidence for patient care. Most patients are excluded from clinical trials, almost systematically, and around 70% of the treatable population wouldn’t be able to participate in a trial. That evidence is what drives the training of LLMs.
When you expect ChatGPT to give an accurate answer but it doesn’t have evidence to train on, where is it supposed to go? That’s anther big place for Atropos. We generate de novo evidence using a chat interface, and we do it based on hundreds of millions of patient records. We’re filling the content gap that exists around this evidence. Not only do we do this every day as we get requests from clinicians and researchers, but we also pre-run thousands of studies that could be fuel for these LLM models to train on.
Lastly, with a forward looking view, with LLMS were largely talking about language models and text. What we’re really excited about at Atropos is – what if you could talk to a database? We’ve seen ChatGPT produce code. Our underlying technology enables us to go from text inputs to database queries to publication-grade responses. So instead of saying “What was the infection rate of covid last month?” and relying on a press release from the CDC, what if you could query the CDC database and get a statistically backed answer? That’s a super interesting answer because now we’re getting into fidelity of data, best data sources to help feed those models…just a couple of areas we are thinking about at this inflection point in clinical Q&A.”