Comment from Sovereign Harbor

AnonymousSupportBusiness
Summary: Sovereign Harbor, a regulatory intelligence and quantitative advisory firm, supports the FDA's initiative to advance real-time clinical trials using AI. They argue that the pilot program must mandate pre-specified decision rules, rigorous statistical frameworks (such as group-sequential designs), and standardized data infrastructure to ensure that real-time data leads to valid clinical decisions rather than acting on noise.
Sovereign Harbor submits the attached comment (PDF) on Docket FDA-2026-N-4390, AI-Enabled Optimization of Early-Phase Clinical Trials Pilot Program. We address questions A.4(b), A.5, B.2, B.3, B.5, and B.6. Central recommendation: the value of real-time data is realised only when the inferential framework is fixed in advance. Continuous access to accumulating trial data without pre-specified decision rules does not accelerate good decisions — it accelerates the risk of acting on noise. We urge the Agency to make pre-specification — of decision rules, error control, estimands, and evaluation metrics — a condition of entry to the pilot. Key points: • Infrastructure (A.4(b)): a capability-based reference architecture — governed data catalogue with full audit lineage (21 CFR Part 11, ALCOA+), open data-sharing standards (no vendor lock-in), privacy-preserving/federated computation, and CDISC/HL7 FHIR standardisation as a precondition of entry. • Timeline & rigour (A.5): pre-specified group-sequential/alpha-spending designs to control Type I error across repeated looks; adaptive features aligned with FDA's 2019 Adaptive Designs guidance; every trial anchored in the ICH E9(R1) estimand framework; real-time safety monitoring kept distinct from confirmatory inference. • Decision quality (B.2): evaluated as diagnostic accuracy and calibration, not speed — concordance beyond chance, predictive value against eventual Phase 2/3 outcomes, and calibration/decision-curve analysis. • Safety & integrity (B.3): time-to-detection and false-alert rate measured against independent adjudication; AI data transformations held to 21 CFR Part 11. • Trustworthiness (B.5): aligned with the NIST AI RMF — external validation (TRIPOD-AI/PROBAST-AI), risk-proportionate explainability, governance evidence without compelling disclosure of proprietary model weights, fairness treated as a statistical-validity problem, and drift monitoring with change-control. • Comparators (B.6): concurrent non-AI controls preferred; where infeasible, externally controlled comparisons adjusted per FDA's 2023 draft guidance; estimand and comparator pre-specified to prevent inflated efficiency claims. The full comment, with footnotes and sources, is attached as a PDF.

View on Regulations.gov