HOSPITALS · HEALTH SYSTEMS · INDUSTRY
Collaborating with the lab
Nearly everything the lab publishes depends on partners in health systems and industry. This page describes how those collaborations typically work.
The shapes a collaboration takes
Data partnerships
A hospital contributes de-identified imaging studies and reports; we handle curation, quality control, and the research built on top. This is how ReXGradient-160K, one of the largest public chest X-ray datasets, came to exist with Gradient Health, and how the MAIDA initiative collects imaging data from hospitals around the world to make medical AI safer and fairer across populations.
Clinical validation studies
A health system's radiologists and clinicians evaluate models on their own patient population: reader studies, site-specific performance evaluations, and prospective assessments of AI-assisted workflows. The partner learns how these systems behave on their patients; the field gets evidence it can trust.
Clinical expertise partnerships
Clinicians co-design benchmarks, annotate and adjudicate model outputs, and shape research questions from real practice. Several of the lab's benchmarks exist because a clinical partner told us what actually goes wrong.
Sponsored research
Industry and health-system partners fund a scoped research question aligned with one of our three programs. Our research has been supported by the NIH, the Biswas Family Foundation in partnership with the Milken Institute, and industry collaborators including Microsoft, Google, and Amazon.
What a partner typically contributes
- Hospitals and health systems: de-identified data under a data use agreement, clinician time for reading or annotation, a clinical question worth answering, and, for validation studies, a setting to evaluate in.
- Industry: funding for a scoped project, and where relevant, compute, data, or engineering collaboration.
What a partner typically receives
- Authorship follows contribution, per standard academic practice: partner clinicians and scientists who shape the work are named collaborators on it.
- Results on their own data and population, not just aggregate benchmarks.
- Visibility into findings as the work develops, within the terms of the collaboration.
The ground rules
- The lab takes on a small number of new collaborations at a time. Work begins with a conversation and a scoped pilot, typically a few months, before anything larger.
- The lab itself does not sign agreements or accept funds. Sponsored research is reviewed and authorized by HMS's Office of Research Administration, industry agreements are negotiated by Harvard's Office of Technology Development, which protects scholarly independence and publication rights, and patient data moves only under executed agreements and appropriate approvals, never through email or web forms.
- We publish. Work that can't eventually be published is usually not a fit.
Starting the conversation
Use the sponsored research form, or come through a lab member who knows your team. A short description of your data or clinical question, your timeline, and what you hope to learn is enough to start. Requests are reviewed as they arrive; email is not part of the intake.