Healthcare has already embraced large language models like ChatGPT and GPT-4. For example, clinicians are using these generative AI tools to assist with clinical notetaking, medical schools are using the technology to play the role of patient for their medical students, and healthcare software giant Epic announced last month its plans to integrate GPT-4 into its electronic health record.
But there hasn’t been a large language model designed specifically for healthcare use — until now.
On Tuesday, Palo Alto-based startup Hippocratic AI emerged from stealth. The company — which bills itself as the first safety-focused large language model designed specifically for healthcare — also gained $50 million in seed funding through a round co-led by General Catalyst and Andreessen Horowitz.
In a recent interview, Hippocratic CEO Munjal Shah said he and General Catalyst CEO Hemant Taneja came up with the idea for the startup while they were on a walk in December.
On that walk, Shah said they discussed how “captivated” they were by ChatGPT and how it was taking over the zeitgeist. They started brainstorming how the new AI model could be used to address healthcare’s massive workforce shortage and burnout crisis.
“We realized there’s really only one way to close that gap in a scalable, fast way because training new folks usually takes years in healthcare. So we said, what if generative AI can really help? In terms of scaling and closing that gap, generative AI can do something that you can’t do any other way,” Shah declared.
From the beginning, Hippocratic’s founding team knew that their AI model would have to take safety incredibly seriously. This is how the startup got its name — it’s a reference to the Hippocratic Oathin which physicians vow to “do no harm.”
With safety in mind, Hippocratic chose to stay away from diagnostics, Shah said. The startup is focused on patient-facing applications, and the large language models that it is developing could be deployed for a number of use cases, he shared.
“What if we gave a free dietician to every person in the country? Think of what that would do to health care costs,” he gave as an example.
Other examples of use cases include a generative AI-powered lactation consultant to help mothers after childbirth or an automated call center assistant who can help people make sense of their medical bills, Shah explained.
“Have you ever seen a healthcare bill that you really understood? We realized we could probably build a much better explanation of benefits and billing customer service agent for every single healthcare call center out there. You’ll call it up and it will remember exactly your file instead of having you tell your story five times while you get transferred from one person to another,” he said.
In Shah’s view, these types of generative AI tools will be helpful because it doesn’t seem like healthcare will ever be able to fill its massive dearth of clinicians with humans. To solve this problem, Hippocratic’s large language models can step in to fill a wide array of clinical roles — such as dieticians, endocrinologists and maternity nurses.
To validate the technology it is developing, the startup plans to mold the AI using human feedback, Shah explained. He said that the Hippocratic is recruiting dieticians, doctors, nurses and other medical professionals to help design the AI models that could one day perform parts of their jobs. These experts will play a key role in determining when each model is ready for testing and eventual FDA reviews, Shah declared.
Since Hippocratic is developing technology that could fill in clinical roles, the startup is already running its AI models through medical role-based examinations, such as certification tests to become a diabetes specialist or lactation consultant.
The startup has run its model through 114 certifications — 106 medical role-based examinations, three standard published benchmarks and five novel bedside manner benchmarks. Its AI outperformed GPT-4 in 105 of these certifications — tying six times and losing three.
To Shah, these early results already communicate the importance of having a generative AI model that is purpose-built for healthcare. The fact that Hippocratic is solely focused on healthcare use cases will help it stand out from commercial generative AI models like GPT-4, he said.
The startup is maintaining its medical focus through its team of healthcare professionals who have been working on developing its AI for several months now — including physicians, hospital administrators, Medicare experts and artificial intelligence researchers from Johns Hopkins, Stanford University, University of Pennsylvania, Washington University in St. Louis, The Health Way, Google and Nvidia.
Photo: Hiron, Getty Images