Stanford researchers tap Oracle Cloud for new ways to treat heart failure

As AI has evolved, we’ve become accustomed to asking virtual assistants such as Apple’s Siri for assistance with both personal and professional matters. What if your doctor could use similar generative AI tools to get proven guidance on treatments that could potentially save your life? 

According to Dr. Cyril Zakka and other researchers at Stanford Medicine’s Hiesinger Lab, that future is sooner than you might think. The lab’s clinicians, engineers, and scientists are using computational protein design, computational fluid dynamics, molecular biology, biomedical devices, and machine learning methods to find new ways to treat heart failure.

Dr. Zakka and his colleagues recently published an article in the New England Journal of Medicine titled Almanac—Retrieval-Augmented Language Models for Clinical Medicine. He calls Almanac’s large language model (LLM) “kind of a ChatGPT for clinical work.” 

All too often, when clinicians ask questions of GenAI tools, the responses are incorrect and the citations are, in Dr. Zakka’s words, “made up.” In contrast, Almanac retrieves information from external databases and unstructured sources, such as the National Institutes of Health’s PubMed and from healthcare data provider BMJ, to answer queries based on “true” data instead of relying on whatever data it was trained on. 

A panel of eight board-certified clinicians and two healthcare practitioners found that Almanac’s responses represented a significant improvement in terms of factuality, completeness, user preference, and adversarial safety compared to the LLMs associated with the most well-known AI tools. 

Need for speed

For demanding generative AI projects such as Almanac, researchers need ready access to high performance NVIDIA GPUs. Dr. Zakka previously experienced this challenge during his medical training at the American University of Beirut, where he founded the Artificial Intelligence in Medicine program, the first of its kind in the Middle East.

With the recent surge in the popularity of AI-related research, the fast GPUs the Hiesinger Lab needed to support the large data sets it uses to iteratively train its algorithms were highly sought after. “We had the funding for new GPUs, but it would have taken four to six months to get them set up,” Dr. Zakka says. “And the GPUs weren’t even available for purchase in the marketplace when we needed them.”

This quest to find high performance GPUs quickly led him to connect with an Oracle team that specializes in helping educational research organizations. Just three weeks later (compared with the aforementioned four to six months), the Hiesinger Lab was running models on Oracle Cloud Infrastructure (OCI) using NVIDIA A10 Tensor Core GPUs.

“We tried working with a few other cloud platforms, but we had to do a lot of the manual setup ourselves,” says Dr. Zakka, a post-doctorate scholar at Stanford. “With Oracle, things went really smooth from Day 1.”

The Hiesinger Lab’s research typically uses models that range from 7 billion to 33 billion parameters. GPUs running on OCI Supercluster and Oracle AI infrastructure with bare metal compute scale up to trillion-parameter models and are optimized for the code that the lab runs. “We could use other GPUs, but we’d have to put a lot more effort into optimizing the code for different hardware,” Dr. Zakka says.

Assisting robotic surgeries

The lab is currently doing background work with OCI GPUs to train algorithms that will advance the use of robotics in surgical simulations. The goal is to generate video footage and images from text across a range of surgical scenarios. As the quality of the simulations improves, the plan is to progress them through a series of test models and regulatory approvals and eventually get them approved for use in surgical training and actual robotic surgeries. 

The main focus of Dr. Zakka’s current research at Hiesinger Lab is to enable its LLMs on OCI GPUs to provide textual interpretations of medical images from procedures such as CT scans, MRIs, and X-rays—to further improve the diagnosis and treatment of cardiothoracic conditions, including for patients awaiting heart transplants. Once he has completed his post-doctoral work, he expects to continue this research and apply it to his medical practice, wherever the future takes him.

“I’m excited about Oracle’s commitment to healthcare,” Dr. Zakka says. “We’re exploring the possibility of getting Almanac deployed in electronic health record systems such as Oracle Cerner EHR. The Oracle team has been super helpful in getting us up and running fast on our research projects. OCI has been an optimal solution, and I expect it will remain that for the people who will be working in my position in the future.”