Building Digital Twins of Immune Responses in Critical Illness: Stanford Receives Funding to Develop a Drug Selection Model and a Clinical AI Oversight System

The immune status of critically ill patients may change before test results become available, making it difficult for physicians to track the course of organ damage in time. Could integrating molecular signals and medical records into a continuously updated model help identify suitable treatments earlier? Newly funded research at Stanford University will explore this approach; another team is addressing a related question: when AI participates in care, who continuously checks its judgments?

Stanford Medicine announced on September 30 that the U.S. Advanced Research Projects Agency for Health (ARPA-H) had awarded research and development contracts to two teams. STEWARD, a project for AI oversight in cardiovascular care, received up to $14.9 million, while IMPRINT, a project for digital twins of immune responses in critical illness, received up to $41.4 million. ARPA-H’s STEWARD page lists up to $15 million, using a more approximate figure; these amounts are development funding, not evidence of treatment efficacy.

Led by Purvesh Khatri, IMPRINT is part of the CIRCLE critical care program. It will measure transcriptomic markers related to the immune system and organs—reading biological activity through gene expression—and combine them with electronic health records to build mechanistic models reflecting the organs affected by critical illness. This “digital twin” aims to simulate changes in a patient’s immune system and organs and suggest potentially suitable drugs already approved by the U.S. Food and Drug Administration (FDA), bringing treatment choices closer to the individual’s biological state.

The model’s predictions will be validated through platforms that combine multiple types of organoids, known as assembloids: different organoids will be assembled to check whether the predictions hold in an experimental system. This can provide biological validation, but it remains a step removed from demonstrating patient benefit. The announcement has not yet specified the size of the Stanford model’s training dataset, its predictive accuracy, or results from validation in humans; existing drug approval also does not mean that new uses or dosing strategies recommended by the model have been endorsed.

CIRCLE also plans for Sage Bionetworks to provide an environment for model development and testing, and for Vanderbilt University Medical Center to help conduct adaptive clinical trials of digital twin-guided immunomodulatory treatment. The overall program has set a goal of reducing intensive care unit stays by 25% and has committed up to $144.9 million over five years, with funding contingent on research milestones. These figures describe the program’s scale and intended targets, not clinical benefits already achieved.

The other project, STEWARD, has Roxana Daneshjou as its principal investigator and is part of the ADVOCATE cardiovascular care program. Its purpose is to oversee autonomous AI agent systems. The team plans three layers of checks: first filtering anomalies, then screening with rules, and finally referring them to a deep research audit agent for verification, generating reviewable reasoning for individual claims. This design progressively increases computational effort, with the aim of enabling clinical teams to receive alerts and trace the evidence behind them.

The two studies place the challenges of clinical AI at different points: IMPRINT must demonstrate that biological signals can support drug selection, while STEWARD must establish a reliable method for continuous AI oversight. The former must bridge the gap between experimental models and real patients, while the latter must address how oversight tools will be evaluated, integrated into care, and meet regulatory requirements. The new contracts provide concrete development paths for these questions, but the answers still depend on subsequent validation.