Evidence-Steered Medicine: A Safety-First Control Logic for Clinical Decision-Making

Background Evidence-based medicine strengthens decision-making, but contemporary care—including clinical artificial intelligence (AI)—often operates under uncertainty, heterogeneous patient contexts, and shifting performance. A common failure mode is committing too early to actions that are difficult to reverse, monitor, or repair. Methods We developed a pragmatic, safety-first control logic by synthesizing concepts from patient safety (including Safety-II), implementation science, and deployment risks in clinical decision support. We operationalized these concepts as a repeatable decision-episode discipline and derived testable hypotheses with pragmatic evaluation designs. Results Evidence-steered medicine (ESM) structures decisions as controlled microsteps: (1) a brief support check, (2) uncertainty banding that constrains action strength, (3) a low-dose action grammar prioritizing reversible micro-interventions paired with short-horizon readouts, and (4) reason-coded governance that enables auditability, learning, and rapid de-escalation/repair. The model yields measurable predictions on severe safety events, recoverability (checkpointing and de-escalation pathways), time to detection of unsafe trajectories, learning efficiency from reason-code distributions, and (in AI workflows) automation-bias and drift-trigger events. Conclusion ESM complements evidence-based medicine by making uncertainty operational: It specifies how to act safely when evidence is incomplete. The hypotheses can be evaluated using retrospective replay, prospective pilots, and stepped-wedge rollouts without replacing standard of care.