From clinical decision support to clinical action systems
What design principles become necessary when medical AI moves beyond classification and begins to change what happens next?
Public research notebook
Thinking in public,
before the field settles.
Arguments, diagrams, open questions, and research hypotheses about medicine after intelligent machines enter the clinical action loop.
Prediction is not participation.
Most medical AI systems end where clinical work begins. They identify a lesion, estimate a risk, summarize a chart, or recommend a next step. A human remains responsible for translating that output into an intervention in the world.
AI-native medicine asks a different design question: what changes when an intelligent system can remain inside the loop—observing the response to an action, adjusting what happens next, and learning across episodes?
The central unit is no longer an input–label pair. It is a clinical episode: state → observation → decision → action → response → outcome, bounded by oversight and the possibility of safe interruption.
Notebook index03 entries
01 Sep 2026
Levels of autonomy in AI-native medicine
A taxonomy for recommendation, supervised action, conditional autonomy, and closed-loop intervention across clinical risk.
What makes a medical action dataset?
Observation, intent, execution, immediate response, outcome, and human correction must be connected as one episode.