Daily English

23 September 2026 · AI × Trials × Validation

When AI enters a clinical trial, what exactly are we validating?

A 30-minute pack on separating AI-as-intervention from AI-for-trial-operations—and deciding how intended use, statistical impact and risk should change validation.

DifficultyC1
Time30 minutes
Main lensintended use
Outputvalidation matrix

Why this reading

“AI in a clinical trial” is not one validation problem.

A new open-access EClinicalMedicine review separates AI-as-intervention from AI-for-trial-operations. The distinction changes what should be prespecified, what evidence matters, and how model changes should be governed.

For SP work, the boundary is especially interesting: an ADaM or TFL agent may be operational infrastructure, but its errors can directly alter inferential results.

Reading order

Your 30-minute plan.

0–3 minPreview

Define the two AI roles.

3–13 minMain review

Trial lifecycle applications and evidence.

13–19 minGovernance

Prespecification, change control, transportability and oversight.

19–24 minICH

E6(R3) risk + E9(R1) estimands.

24–25 minSP bridge

Classify three AI tools.

25–30 minOutput

Build a TFL-agent validation matrix.

Open-access sources

A current review plus two regulatory lenses.

Main · EClinicalMedicine · 12 Sep 2026 · Open accessArtificial intelligence in clinical trials—state of the evidence, gaps, and next stepsAI roles, evidence, agentic workflows and governance ↗Support · EMA / ICHICH E6(R3) Good Clinical PracticeCritical-to-quality factors and proportionate risk ↗Support · FDA / ICHICH E9(R1): Estimands and Sensitivity AnalysisDefine the treatment effect before analysis ↗

Brief background

Start with intended use, not the model name.

An AI recommendation assigned to a participant, a recruitment screener, and an ADaM-code generator all use AI, but they occupy different positions in the evidence chain. Validation should therefore depend on what the system can change.

AI-as-intervention needs prospective clarity around model version, inputs, change control and the treatment effect of interest. AI-for-trial-operations still needs evidence, but the validation target depends on consequence.

ICH E6(R3) gives a risk lens: participant protection, reliability of trial results, detectability of failure, and critical-to-quality factors. E9(R1) gives a statistical lens: a program is only correct if it implements the treatment effect the trial actually intends to estimate.

For an SP agent, preserve intended use → inputs → version → output → decision influenced → failure mode → controls → human review → audit trail.

Key vocabulary

Fifteen terms for AI validation in trials.

Term中文Meaning
AI-as-interventionAI作为干预AI whose output is part of the assigned clinical intervention.
AI-for-trial-operationsAI用于试验运营AI supporting trial design, conduct, analysis, or reporting.
estimand估计目标The treatment effect a trial aims to estimate.
prespecification预先规定Defining objectives and methods before outcomes are known.
change control变更控制Governed review, approval and versioning of changes.
transportability可迁移性Validity of performance in another population or setting.
endpoint adjudication终点评定Structured determination of whether an endpoint occurred.
inferential modelling推断建模Statistical modelling used to draw treatment-effect conclusions.
critical-to-quality factor关键质量因素A feature fundamental to participant protection or reliable results.
risk-proportionate风险相称的Matching controls to meaningful risks.
model drift模型漂移Performance change as data or environments evolve.
auditability可审计性Ability to reconstruct actions, versions and approvals.
human oversight人工监督Qualified human review or decision authority.
workflow accuracy工作流准确性Reliability of an AI-supported operational process.
decision impact决策影响The consequence if an AI-supported decision is wrong.

Useful phrases

Language for validation and governance discussions.

  1. the validation target depends on the role AI plays in the trial.
  2. an operational tool can still become inferentially important.
  3. prespecification limits hidden flexibility after outcomes are known.
  4. change control should cover both model versions and surrounding workflow logic.
  5. performance must be evaluated in the population and setting where the system will be used.
  6. risk-based oversight should focus on critical-to-quality factors.
  7. a high-accuracy component may still create unacceptable downstream risk.
  8. the same AI system may require different evidence under different intended uses.
  9. traceability should preserve inputs, versions, decisions, and human approvals.
  10. automation should be judged by consequence as well as technical performance.

Comprehension

Five questions.

  1. What is the central difference between the two AI roles?
  2. Why can operational AI still become inferentially important?
  3. How does E6(R3) help scale validation effort?
  4. Why can correct SAS still implement the wrong estimand?
  5. What evidence makes an AI-generated TFL workflow auditable?

Retelling

Say it three times.

  • 30 seconds · Explain the two AI roles without saying treatment or operations.
  • 45 seconds · Endpoint classifier vs email assistant: why governance differs.
  • 60 seconds · Intended use → decision impact → controls → human review.

5-minute output task

Build a validation matrix for an AI-assisted TFL agent.

  1. Minute 1: Define intended use and prohibited autonomy.
  2. Minute 2: Name three result-changing failure modes.
  3. Minute 3: Add deterministic controls.
  4. Minute 4: Define audit evidence.
  5. Minute 5: State the mandatory human-review boundary.

One sentence to keep

The right question is not whether AI is accurate in general, but whether its evidence, controls, and oversight are proportionate to the decision it can change.