The Archive

Introducing the five-component framework for population-level algorithmic surveillance in clinical medicine.

The Novartis settlement — 75 years later. The same extraction, a new medium, and a legal theory that is very much alive.

AI doesn't treat patients directly. It changes how clinicians think. That distinction is the foundation of the entire discipline.

Defining dose, frequency, and population at risk for clinical AI systems — the vocabulary the field is missing.

The FDA cleared over 700 AI/ML-based medical devices. Post-market surveillance requirements are minimal. Here is what that actually means.

A method built for exactly this problem — and almost never applied to clinical AI deployment.

A practical protocol for detecting when clinicians have stopped questioning the algorithm — and started deferring to it.

Introducing the AAECS — the Algorithm Adverse Event Classification System — and the case for mandatory reporting.

The most underused trial design in implementation science — and why it maps precisely onto phased algorithm rollout.

The methods that catch drug safety signals years after approval — adapted for clinical AI post-deployment monitoring.

What a properly designed registry requires — and why registry design is epidemiological architecture, not just logistics.

Alert fatigue is not a human failure. It is an algorithmic design failure with population-level consequences.

A device in use for decades. A bias documented in the literature. A surveillance failure that compounded across a pandemic.

Training data bias in dermatology AI — and what it takes to build surveillance that catches differential performance before it becomes differential harm.

The algorithm flagged a patient at 2am. The nurse overrode it. Three hours later, the patient coded. Who failed?

When AI assumes the interpretive work, the question changes. "What do I see?" becomes "Do I agree?" Those are not the same cognitive act.

One of the most widely deployed clinical AI systems in the world. What we know, what we don't, and what a proper surveillance study would look like.

A precise legal and ethical analysis of liability in algorithm-mediated clinical harm. The answer is more complicated — and more urgent — than you think.

Every hospital has a CMO and a CIO. Neither is responsible for what happens when their domains collide.

The most ambitious AI regulatory framework read through an epidemiological lens. A clinician's guide to what's coming.

If every drug has an NDC and every device has a UDI, every clinical AI system needs a CAIT. Here is what that would require.

A draft institutional policy any hospital system could adapt — covering procurement, deployment, monitoring, and incident response.

Peer review, editorial standards, and the infrastructure a new discipline needs to produce credible science.

What medical students and residents need to know about clinical AI — and why they are not learning it now.

You approved these systems. You are responsible for what they do after deployment. Here is what that responsibility requires.

One year. Twenty-seven issues. The argument is made. Now we build the institution.