Chip giant Nvidia is collaborating with startup Abridge to train healthcare-specific foundational models tailored to clinical conversations.
The AI models are designed to improve the accuracy, reliability, auditability and customizability of clinical workflows, from documentation and evidence base to workflow automation and clinical reasoning support, the companies announced Thursday.
Based on the Nvidia Nemotron open model family, where both model weights and training data are available, the new Abridge models will be trained on Nvidia Blackwell AI infrastructure using advanced pre-, during-, and post-training processes using anonymized data, the companies said.
Abridge executives noted that training across all three stages enables embedding of clinical knowledge from the ground up, improving accuracy, precision and reliability across specialty areas, clinical settings and multi-step workflows following clinical conversations. By adapting domains early in the training lifecycle, Abridge can build models that make clinical inferences from the ground up. Nvidia’s Nemotron enables Abridge to optimize quality, cost and efficiency at every layer and deploy the right models for the right workflows at the right scale, executives said.
Nvidia is also an investor in Abridge through its NVentures venture capital arm.
Abridge is rapidly building an AI platform that goes beyond just an AI scribe tool and functions as a full-fledged AI clinical assistant. This week, the company announced a major platform expansion to integrate payer and life sciences workflows. Described as an “AI-native clinician intelligence platform,” Abridge says it now connects care delivery, payment, and evidence-based treatment.
The company currently works with 300 health systems and its technology supports more than 100 million conversations annually. As Abridge builds its AI platform, we want to improve the performance and speed of our models.
“You want to sprinkle intelligence everywhere, but how do you find a way to do that? And how do you also find a way to provide the right level of accuracy and latency? Ultimately, you’re going to need a little more control in some areas than you expected, but we’re very proud to work together to build on the Nemotron model family,” said Abridge CEO. and co-founder Shiv Rao, MD, spoke on stage Thursday at a keynote event in New York City. With Kimberly Powell, Vice President of Healthcare, Nvidia.
Dr. Rao continued, “We have a very maximalist attitude of sprinkling intelligence everywhere we can, which means we need to reach down to the bottom of the stack and control our own destiny. Latency is a big issue for us, because we want clinicians to be able to stop, rotate their chairs, and put all the artifacts there.”
Health tech companies see a huge opportunity to leverage Nvidia’s computing power to advance their AI capabilities. Last fall, Verily, part of Alphabet and Google’s life sciences sister company, announced a partnership with Nvidia to integrate the company’s AI technology stack into Verily’s Pre platform. Innovaccer also announced that it has adopted Nvidia’s full-stack AI platform to accelerate voice, text, and multimodal inference infrastructure to power its AI agents.
The computer chip giant is also moving deeper into life sciences. Eli Lilly and Roche have entered into an AI infrastructure partnership with Nvidia, Fierce Biotech reported. On the medical technology front, the company is working with Thermo Fisher Scientific to build autonomous laboratory infrastructure and with Netherlands-based diagnostics manufacturer Qiagen to increase researchers’ ability to leverage AI in the drug discovery process, as reported by Fierce Medtech. The company is also moving deeper into medical technology and cancer research through a partnership with diagnostics maker Droplet Biosciences.
Powell said Thursday that Nvidia is bullish on health care “going to be one of the largest technology industries.”
“NVIDIA is not, never will be, and never will be a healthcare company, but we believe we have the unique ability to contribute to tackling some of the world’s toughest and most impactful jobs,” she said.
“We are in a phase of rapid evolution in AI, so how can we make it our mission statement to make this rapidly evolving technology available to the healthcare industry?” she said.
Powell noted that the tech industry has evolved through “three major technological breakthroughs” in the past 18 months, from “AI that can create things, to AI that can reason about things, to AI that can do jobs.”
He noted that common AI models “don’t understand clinical language, don’t have clinical reasoning, and certainly don’t have the expertise for all the long-running tasks and interconnected work needed to completely transform workflows, so we need AI intelligence built for more domain-specific, complex workflows in healthcare.”
“Any technological breakthrough requires a full-stack focus. We think of AI as a full-stop computing problem, not a model. It starts with energy, then chips and systems, then AI data centers and cloud, then core foundation models, and finally the application layer.[Abridge]made the case head-on that there is incredible value and effectiveness in the vertical AI application layer,” Powell said Thursday. “What we recognize together is that it’s time to dig deeper into the stack, the clinical conversation infrastructure model, to enable the complexity of healthcare, all the workflows, and the connectivity of this amazing ecosystem that you’ve enabled, because it needs to become more domain-specific.”
