Mayo Clinic taps Microsoft and Cerebras for clinical AI models

Getting a clear medical diagnosis often comes down to two bottlenecks: radiologists drowned in image backlogs, and patients enduring months of painful trial-and-error just to find the right medication.

Mayo Clinic wants to tackle both sides at once. At the 43rd Annual J.P. Morgan Healthcare Conference, the health system unveiled separate collaborations with Microsoft Research and Cerebras Systems to build specialized generative AI foundation models for radiology and genomics.

AI assistants for the radiology room

Mayo's project with Microsoft Research focuses on chest X-rays, the workhorse of hospital diagnostic imaging.

The teams are developing multimodal foundation models trained on Microsoft's AI technology and Mayo Clinic's clinical imaging data, including X-rays alongside CT scans and MRIs. The goal is to build proof-of-concept systems that can automatically generate draft radiology reports, verify the placement of tubes and lines, and highlight critical changes compared to prior patient scans.

Dr. Matthew Callstrom, medical director for Generative AI and Strategy at Mayo Clinic, noted that multimodal foundation models could break longstanding bottlenecks across radiology workflows by giving clinicians instant context.

Jonathan Carlson, managing director of Health Futures at Microsoft Research, said the tie-up aims to turn biomedical research into practical tools that clinicians can rely on for more accurate care.

Cutting trial-and-error in drug prescriptions

Mayo's partnership with AI chipmaker Cerebras takes aim at personalized medicine through DNA.

The two created a genomic foundation model combining public human reference genome data with patient exome data from the Mayo Clinic Platform. Exome sequencing zooms in on protein-coding regions, where the vast majority of disease-causing mutations reside. Cerebras provides the wafer-scale hardware necessary to train models on these sprawling biological sequences.

The resulting model compares patient genomic profiles with similar historical cases in near real-time to forecast disease trajectory and drug efficacy. In conditions like rheumatoid arthritis, patients typically spend months cycling through costly immunotherapies before finding one that induces remission. Mayo reported that early benchmark tests show promising accuracy in predicting patient treatment responses early on.

Why it matters

General-purpose chatbots sound impressive discussing symptoms, but clinical medicine requires specialized foundation models anchored in raw medical pixels and genetic code. Mayo Clinic's parallel plays highlight the blueprint for hospital-grade AI: narrow the operational scope and partner with domain specialists.

We think Mayo's sequencing choice is particularly smart. Focusing on exome data rather than full genomes cuts computational waste while concentrating on actionable disease mutations. Meanwhile, turning to Cerebras underscores that genomic sequence modeling requires massive context windows and memory bandwidth, opening doors for alternative silicon outside Nvidia's walled garden.

Still, the clinical stakes remain unforgiving. Spotting misplaced lines on a chest X-ray or steering autoimmune prescriptions carries zero margin for hallucinations. Mayo's patient exome pool also skews toward its own demographic base, meaning these models must prove their real-world transferability across diverse patient populations before doctors should trust them with treatment calls.