AI in clinical trials.
The honest map of AI in trial conduct sorts use cases by what they actually replace and how mature the regulator's posture is. Some are production, some inspection-acceptable with caveats, some piloting, one is regulator-flagged as high-risk.
AI in clinical trials.
Use cases · what each replaces or augments · maturityThe honest map of AI in trial conduct sorts use cases by what they actually replace and how mature the regulator's posture is. Some are production (deployed at sponsors and inspection-acceptable), some inspection-acceptable with caveats, some still piloting. One is regulator-flagged as a high-risk finding.
Cohort enrichment · site selection.
Patient matching to inclusion criteria; demographic targeting; geographic optimisation. Replaces manual eligibility screening; augments site selection. Standard 2024+ at Medidata, IQVIA, Syneos. Inspection risk low — site selection aid, not protocol violation.
Real-time eligibility screening at site.
Augments inclusion/exclusion documentation accuracy. Emerging 2024–2026, not yet standard. Regulator gates on protocol fidelity, not AI mechanism. Risk: if AI screen misses an exclusion, FDA cites protocol deviation — not the model.
Adaptive design optimisation.
AI-assisted sample-size recalculation; dose-escalation sequencing; interim-analysis decisions. ICH E20 traditional framework mature; AI-extensions under discussion. PCCP framework (device-side) may extend conceptually. Adaptive design already complex; AI adds a layer.
AI-assisted SDV.
Augments manual record review. Vendors piloting 2024+. No specific framework yet. Risk medium — overreliance on algorithm vs. auditor judgment. Inspectable when paired with risk-based monitoring SOP.
Safety signal detection.
AE signal identification across pharmacovigilance datasets. Deployed in some CROs and sponsors. Sits in PV regulation (E2A) rather than CT regulation. Risk low — supplementary to human review.
Synthetic control arm curation.
AI + real-world data for external control cohort selection. FDA-led rare-disease and oncology pilots. SCA frameworks developing; AI-specific mechanism underspecified. Risk medium — representativeness contested, data quality.
Informed consent simplification (NLP).
ICF readability assessment. Research-stage. No specific framework. Risk low if used for readability check only, not authorship.
DCT logistics optimisation.
Telemedicine visit scheduling; direct-to-participant IMP shipment routing. Operational tool, not protocol-affecting. Emerging 2024–2026. Risk low.
Generative AI authoring · protocols, ICF, CSR.
Replaces human writing — if allowed. Proposed by some vendors. FDA, EMA, PMDA have signalled unacceptable without human-in-the-loop. High-risk inspection finding if detected. Authorship integrity question is not solved.