Dramatically accelerating patient-matching in clinical trials with open source

The open source AI landscape is transforming the way organizations like Mendel AI, a leading clinical AI platform, access cutting-edge technology and drive social impact. By leveraging models like Llama, companies can create tailored AI solutions for specific fields without the hefty upfront costs associated with proprietary systems or the need to share their data with the model’s parent company—an important factor in highly regulated industries, like clinical trials.

The company’s flagship product, Mendel Hypercube, tackles clinical tasks like data abstraction, chart reviews, and patient cohort analysis. With Llama, Hypercube can also be used to chat and query in natural language for trial matching and putting together patient groups. Studies have shown that matching patients with clinical trials can take hundreds of days, leading about 80% of clinical trials to miss enrollment goals, but Hypercube can accomplish this in just one day.

Hypercube combines Llama with a clinical hypergraph. The platform allows healthcare companies to organize their data on their own cloud, creating a secure and searchable knowledge base.

“We’re about to make a huge leap in patient outcomes,” says Dr. Wael Salloum, Mendel’s Founder and Chief Science Officer. “Open source models help companies innovate quicker by letting them concentrate on their specific needs and add advanced reasoning capabilities to their AI stacks.”

Mendel’s introduction to Llama began by fine-tuning a language user interface task with Llama 2, allowing users to communicate with Mendel’s knowledge base inference engine, translating natural language questions into its symbolic query language.

To create its healthcare-specific foundation LLM, the Mendel team continuously pre-trained Llama 3, using both 8B and 70B.

“We have a variety of LLM-based healthcare tasks, and we continuously do supervised fine-tuning of our foundation LLM on labeled data for them,” says Salloum.

The result is a collection of lightweight models with instruction-following and in-context learning capabilities in an agentic framework, which can abstract patient records and answer research questions on large databases of patients.

“As foundation LLMs become more of a commodity, having an open source version that can save startups the initial development costs is a huge contribution to society,” Salloum says. “Innovators can spend their time thinking about building on top of such models by figuring out customer use cases and other non-LLM reasoning and cognitive capacities to add to their AI stacks.”

Looking ahead, Salloum notes that Mendel AI is planning to use the multimodal Llama 3.2 in the near future.