MedRise MedicineAI in medicineWhere AI is already used in patient care, how the FDA treats it, what the evidence shows when tools are tested outside the lab, how to use AI safely while you train, and the real routes into clinical informatics and health AI research, with the conferences and deadlines worth knowing.
The landscape
"AI" covers very different tools. Knowing which kind you're looking at tells you how it was tested and who is responsible for it.
Regulation
The FDA's own position is that it does not regulate AI as such; it regulates medical devices, including AI-enabled ones. So the question is always what a product is intended to do. Software that analyzes an image to detect disease is a device; much of the AI used for scheduling, billing or general writing is not.
The FDA says it had authorized over 1,600 AI-enabled medical devices for the U.S. market as of September 2026, and it cautions that its public list is not a complete record of every AI-enabled device. An analysis of the list found about 96% reached the market through the 510(k) pathway, which compares a device with one already on the market.
In December 2024 the FDA finalized guidance on predetermined change control plans. A maker can describe, at the time of review, the model updates it plans and how it will test them, then make those changes without a new submission each time. The final version covers all AI-enabled device software, not just machine learning.
A cleared device was shown to work for a specific intended use, population and setting. It does not mean it works for every patient, at every hospital, or for questions it was never cleared to answer. Read the intended use before you trust the output.
Evidence
Two studies every trainee should know, because they show the two most common failure modes.
Researchers at Michigan Medicine tested a widely used proprietary sepsis model on 38,455 hospitalizations. It reached an area under the curve of 0.63, missed 67% of patients who developed sepsis, and still alerted on 18% of all hospitalized patients. Lesson: a model can be deployed widely without strong external validation, and poor accuracy also creates alert fatigue.
A commercial algorithm used to pick patients for extra care predicted future health care costs rather than illness. Because less money is spent on Black patients at the same level of need, Black patients were considerably sicker than White patients at the same risk score. Fixing the bias would have raised the share of Black patients flagged for extra help from 17.7% to 46.5%. Lesson: check what the model actually predicts, not what it is called.
Critical appraisal
Medical education
The AAMC has published Principles for the Responsible Use of Artificial Intelligence in and for Medical Education (seven principles, released January 2025), with workshop materials for faculty, and separate principles for AI in medical school and residency selection, which say any use of AI in selection should be balanced with human judgment. Expect your school to have its own policy built on them.
Careers
Clinical informatics became a subspecialty open to diplomates of all ABMS member boards in 2010, and the first physicians were certified in 2014. Certification is offered jointly by the American Board of Preventive Medicine and the American Board of Pathology. Today the only route to the initial exam is a 24-month ACGME-accredited clinical informatics fellowship after residency: the practice pathway closed after the 2025 application cycle.
Start now
MIMIC and many other de-identified clinical datasets need PhysioNet credentialing: an account, your affiliation, the CITI course “Data or Specimens Only Research” (upload the completion report, not just the certificate) and a data use agreement for each dataset. PhysioNet says approval typically takes 1 to 2 weeks.
Testing how an existing, published score or model performs in a new population is publishable, teaches you the methods, and answers the question clinicians actually care about.
Biostatistics (sensitivity, calibration, ROC curves), a little Python or R, and how EHR data are recorded. These outlast any single tool.
AMIA for clinical informatics, SIIM for imaging informatics, and the machine learning for health research community (ML4H and CHIL). Ask each society about student membership and student registration rates.
Imaging and applied informatics projects fit SIIM's abstract calls; AMIA's spring and fall meetings take informatics research and student papers. Deadlines are listed below.
Conferences
From the MedRise conference list, which is re-checked against each organizer's own page every week.
MedRise helps students and IMGs choose a project, get it published and present it at the right meeting.
Sources
Sources last checked November 14, 2026. Device counts, guidance and certification rules change: open the official page before you rely on a detail.