AI models in healthcare have shown strong predictive performance, which has not translated into better care or lower cost — a gap often called the AI chasm. The value of using an AI model is determined by the action pairing it enables: the lead time of the model's prediction or recommendation, whether an effective intervention exists, the work capacity to deliver it, and whether the resulting allocation of resources is fair. The focus, then, is not about building a better model. It is delivery science: making models workflow-aware, evaluating usefulness and fairness during development rather than after deployment, and bringing AI to the clinic safely, ethically, and cost-effectively. There are three efforts aimed at this goal.

The academic lab, which is a mix of doctors, engineers, informatics professionals, and students developing methods to learn from patient-level health data, answer clinical questions at the point of care, and research safe, ethical, and cost-effective use of predictive models. It is affiliated with the Department of Medicine, the Clinical Excellence Research Center, and the Department of Biomedical Data Science. The current portfolio is organized around five themes, anchored on live systems and pairing research methodology with a real deployment to study: (1) evaluation infrastructure for clinical AI, including MedHELM and HealthAdminBench; (2) safety, reliability, and guardrails for generative clinical AI; (3) post-deployment learning from ChatEHR; (4) EHR foundation models and long-context representation; (5) rare disease phenotyping and discovery. Prior work at shahlab.stanford.edu/examples_of_prior_work and shahlab.stanford.edu/rail