Stanford University
Schmiedmayer LabDivision of Computational Medicine

AI Agents for Patient-Centered Care, Prevention, and Biomedical Research.

We develop computational methods for multimodal AI agents that retrieve and interpret evidence from wearable signals, electronic health records, and patient-reported experiences. We study how these agents can help patients and care teams understand changes in health between visits, and how AI co-scientists can help researchers test hypotheses across large biomedical datasets.

Explore the researchMeet the team
Editorial illustration of a person using a phone and wearable outside the clinic, connected through layers of health information to a clinician reviewing records.

Research

Multimodal Computational Methods

We develop methods to locate relevant intervals in physiological recordings and relate them to symptoms, activity, and care events without reducing a long history to a single average.

Models and benchmarks

Learning and Evaluation from Patient Interactions

We use portal messages, interviews, and behavioral studies to identify what patients need between visits and to evaluate whether an agent’s explanations address those needs.

Studies and evaluation

Patient-Centered Agentic Systems

We study how agents can support patients over time, when a clinician should review a proposed action, and how researchers can verify an agent’s analysis against its data and code.

Agent supervision

Computational medicine

Reasoning across wearable signals and clinical records.

Our methods research focuses on representations that let AI reason across physiological time series, clinical records, and patient-generated data. We investigate how to preserve temporal structure, learn from incomplete observations, and use long health histories efficiently.

OpenTSLM takes time-series signals as input alongside text. To teach models about patterns that clinical notes rarely describe, TimeCap generates captions directly from wearable recordings. With PocketTSLM, we study how much of that reasoning can happen on a patient’s phone.

Editorial collage of a wearable, clinical records, and a phone and journal representing patient context.
  • Time seriesWearable and physiological signals
  • Health recordsClinical notes and measurements
  • Patient contextExperiences and questions

Patient-facing AI in practice

Patient-facing agents for longitudinal care.

Our work on cardiometabolic prevention starts with the questions patients send their care teams about symptoms, test results, and medications. In spine surgery and reproductive endocrinology and infertility (REI), interviews with patients and clinicians guide what an agent should explain before a visit.

Plainly

In consented studies, patients can explore their health records through chat or voice. We evaluate how well an explanation reflects the record and whether it helps patients prepare for a conversation with their care team.

Patient-message researchSpine surgery and REI studies
Plainly study chat with a synthetic health-record explanation and a participant question.Plainly voice demonstration reading a synthetic health-record explanation aloud.

Understanding everyday health.

Phones and wearables can help us study health outside the clinic. They also shape how we spend our time. We investigate both their role in prevention and the challenges of social-media and smartphone use.

Learning from daily life

In our My Heart Counts collaboration, wearable measurements and participants' experiences help us study cardiovascular health and physical-activity coaching. These studies also inform how we develop and evaluate multimodal models.

Digital wellbeing and agency

We study patterns of smartphone use, problematic social-media use, and interventions that help people make their own choices. Our work examines what app activity reveals beyond screen time and how to evaluate changes in behavior and wellbeing.

AI co-scientists for biomedical research.

We study how agents can help investigators test whether findings hold across cohorts, devices, and clinical settings. In Heartwood, the investigator can review the proposed analysis before it runs and trace each result back to the data and code.

Heartwood action review showing proposed commands for a synthetic cohort analysis.

Researcher review in Heartwood

  1. Plan an analysis

    Turn a scientific question into a plan a researcher can review.

  2. Work with data and code

    Use tools to inspect data, run analyses, and investigate errors.

  3. Review the evidence

    Keep methods, outputs, and decisions available for others to check.

Editorial illustration of researchers reviewing code and sharing methods around a common worktable.

Open source is part of how we do research.

We share models, code, and research tools so other teams can inspect, adapt, and extend our work. We are developing Grove and Trove to make building applications and running studies easier.

Grove
A modular framework for building digital health applications.
Trove
An open-source study platform in beta for research teams and participants.
DIGITS
A Harvard and Stanford collaboration on shared standards for wearable data.

We welcome code contributions, documentation, reproducible examples, and feedback from patients, clinicians, and researchers.

Find a way to contribute

Work with us

We welcome students, researchers, engineers, and collaborators interested in multimodal models, patient-facing AI, digital wellbeing, and biomedical research agents.

Join the labContact the lab