arXiv:2609.00076 Defines AI M&M Four-Axis Classification for Clinical Failures

The paper formalizes case intake, evidence preservation, tool-in-loop attribution, and corrective-action tracking to reconstruct AI-in-workflow errors. It separates the exposing condition from the risk-producing process, the care consequence, and the assigned remediation. Two independent reviewers applied the axes to medication and decision-support examples without disagreement.

Existing aggregate monitoring detects performance drift while traditional safety reports capture events, yet neither reconstructs interactions among models, clinicians, workflows, and controls. The framework converts isolated failures into institutional learning without replacing regulatory oversight or statistical surveillance. It draws on classic M&M conference structure but adds explicit attribution for AI components.

Analysis shows the four-axis separation enables targeted fixes rather than generic retraining. Full agreement on 20 classifications in the illustrative set indicates initial reproducibility, though the paper notes this remains untested at scale. Prospective multi-institution deployment will determine whether the method surfaces systemic vulnerabilities missed by current post-market surveillance.

Next steps require documented pilots that track time-to-remediation and recurrence rates after AI M&M reviews. Without those data, claims of improved patient safety rest on structural logic alone.