Regulation — pharma-specific

FDA AI/ML SaMD Action Plan — GMLP + TPLC for AI in clinical trials

The Action Plan (2021, updates 2024-2026) reframed AI as a lifecycle question, not a snapshot approval. Good Machine Learning Practice (GMLP) principles and Total Product Lifecycle (TPLC) stance apply to any AI that touches a regulated trial — not just AI in cleared medical devices.

What the action plan actually says

"FDA envisions a Total Product Lifecycle (TPLC) approach to AI/ML-based SaMD … that would allow for evaluation and monitoring of a software product from its premarket development to postmarket performance along with the real-world performance of the algorithm." FDA Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan (January 2021)

"Good Machine Learning Practice (GMLP) … to help promote safe, effective, and high-quality medical devices that use artificial intelligence and machine learning." FDA Action Plan — guiding principle #2 (updated 2024 GMLP principles jointly with Health Canada + MHRA)

What this means in plain English

Three practical stances the SaMD Action Plan pushes toward:

Lifecycle, not launch. A pre-market snapshot is not enough. Real-world performance monitoring runs the life of the AI.
GMLP as evidence. Data-management practices, training-data representativeness, and monitoring discipline are all reviewable — not just model output.
Traceable modifications. Changes to the AI over its life must be traceable to a rationale + evaluation. Silent retrainings do not satisfy this.

For a trial-eligibility classifier, this means the sponsor cannot delegate "the model works fine" claims to the clinical-AI vendor's uptime dashboard.

What triggers the exposure in the sample

The trial-eligibility AI ran the full 90-day audit period without retraining — while P3 KL-divergence hit 1.126 (5x the flag threshold) on Day 51. That is exactly the "real-world performance drift" TPLC is calibrated to detect. The GMLP question — "does the sponsor have monitoring practices sufficient to catch this class of drift?" — now has a hard-to-answer shape without independent evidence.

What the $499 Snapshot shows against this rule

Real-world drift-event log — the TPLC-monitoring evidence FDA expects
KL-divergence + variance-ratio + mean-shift thresholds documented + timestamped
Model-version pinning per decision — traceable-modifications evidence
Independent-verifier signature — separates the monitoring plane from the model being monitored

See the drift-detection chart →

How does this help me?

The Action Plan is a direction, not a rule. But CDRH + CDER reviewers apply its GMLP principles when asking sponsors about AI-in-trial evidence. Having independent monitoring on file shifts that conversation before it starts.

Read: FDA AI/ML SaMD — what it saves you in the GMLP conversation →

$499 Snapshot. 3 business days.

The independent-monitoring record GMLP principles ask for — on your actual trial-AI surface.

Buy $499

Snapshot credit applies to Baseline ($2,500) or Enterprise Attestation ($35-55K) upgrade within 30 days.