Granular Storage For Future Workflow Analysis

We wish to capture enough structured history to support traceability, QA, debugging, and future analysis of how radiologists use AI suggestions.

Audit Points

Moment	Stored event	Future analysis value
Study received	Who/what sent it, when, metadata digest, validation status.	Integration reliability, missing metadata patterns.
AI run completed	Model version, inputs, outputs, confidence distribution, latency.	Model monitoring, drift analysis, per-hospital behavior.
Radiologist views study	Viewer opened, overlays toggled, AI finding selected.	Adoption, trust, attention, workflow friction.
Finding feedback	Accepted, rejected, edited, ignored, reclassified, or superseded.	Human-AI agreement/disagreement analysis and evaluation.
Legacy report completed	Completion timestamp, report reference if available, linked sidecar findings/annotations.	Outcome correlation and sidecar workflow evaluation.
Status / handoff recorded	Destination or source system, success/failure, retry count, payload digest if a handoff exists.	Operational support and hospital onboarding checks without assuming we own final submission.

Which AI findings are most often accepted without edits?
Which finding types are frequently resized, relabeled, or rejected?
Do radiologists review AI suggestions before or after reading the raw image stack?
How often does the final diagnosis include findings not suggested by AI?
Where do turnaround delays occur: ingestion, inference, assignment, sidecar review, or legacy report completion?
Do different hospitals have different disagreement patterns after normalizing for modality and model version?

Compliance Boundary Production implementation would require a dedicated security, privacy, and regulatory review.