Why Science? | ICH E20 Adaptive Designs for Clinical Trials

The ICH draft was endorsed at Step 2b on 25 June 2025 and moved into public consultation. Its focus is confirmatory trials intended to establish efficacy and support benefit-risk assessment, though the principles can inform other phases. Draft recommendations do not supersede current regional guidance until finalisation and implementation.

The beauty of e20 is a step 3 draft is disciplined learning. The ICH draft was endorsed at Step 2b on 25 June 2025 and moved into public consultation. Its focus is confirmatory trials intended to establish efficacy and support benefit-risk assessment, though the principles can inform other phases. Draft recommendations do not supersede current regional guidance until finalisation and implementation. Keep this boundary visible: E20 remained a draft in Step 3 on 27 October 2026.

A strong science note separates observation from interpretation. Trace e20 is a step 3 draft from interim data cut through statistic, threshold, authorised decision and final inference. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Next chapter

The defining feature is a prospectively planned opportunity to modify specified aspects of an ongoing trial based on interim information. The modification can be rule based or use a bounded decision process. Changes unrelated to comparative interim data, such as ordinary operational adjustments, are not automatically adaptive-design features.

For an adaptive design changes a planned feature using accumulating data, every branch belongs to one prospective strategy. The defining feature is a prospectively planned opportunity to modify specified aspects of an ongoing trial based on interim information. The modification can be rule based or use a bounded decision process. Changes unrelated to comparative interim data, such as ordinary operational adjustments, are not automatically adaptive-design features. Preserve this guardrail: Prospective adaptation is not permission to redesign a trial after comparative results are known.

A strong science note separates observation from interpretation. Challenge the timing, delayed outcomes, error control, estimation and operational firewall behind an adaptive design changes a planned feature using accumulating data. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A confirmatory trial should adapt only when the change improves efficiency, ethics, feasibility or information without weakening the objective. The rationale connects the adaptation to the development programme. Complexity that cannot affect a meaningful decision adds operational and interpretive risk.

A trustworthy account of every adaptation needs a scientific reason connects interim information, action and final inference. A confirmatory trial should adapt only when the change improves efficiency, ethics, feasibility or information without weakening the objective. The rationale connects the adaptation to the development programme. Complexity that cannot affect a meaningful decision adds operational and interpretive risk. Remember: Type I error control does not by itself guarantee unbiased estimation or operational integrity.

A strong science note separates observation from interpretation. Reconstruct every adaptation needs a scientific reason with protocol version, code, random seed, interim report and decision minutes. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Population, treatment conditions, outcome, intercurrent-event strategy and summary measure define the treatment effect of interest. An adaptation can change enrolment or allocation while the estimand remains stable, or it can create a different target population. The analysis must stay aligned with the effect the trial claims to estimate.

The beauty of the trial still answers an estimand is disciplined learning. Population, treatment conditions, outcome, intercurrent-event strategy and summary measure define the treatment effect of interest. An adaptation can change enrolment or allocation while the estimand remains stable, or it can create a different target population. The analysis must stay aligned with the effect the trial claims to estimate. Keep this boundary visible: Simulation evidence is bounded by the scenarios, models and implementation code that were tested.

A strong science note separates observation from interpretation. Test the trial still answers an estimand under null, target, heterogeneous, missing-data and time-trend scenarios. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

The protocol and statistical plan specify the number and timing of interim looks, which data are used, how maturity is assessed and what actions are available. Calendar dates alone may be misleading when the design depends on event counts, follow-up or information fraction.

For interim analyses need planned timing and information, every branch belongs to one prospective strategy. The protocol and statistical plan specify the number and timing of interim looks, which data are used, how maturity is assessed and what actions are available. Calendar dates alone may be misleading when the design depends on event counts, follow-up or information fraction. Preserve this guardrail: E20 remained a draft in Step 3 on 27 October 2026.

A strong science note separates observation from interpretation. Trace interim analyses need planned timing and information from interim data cut through statistic, threshold, authorised decision and final inference. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A rule may stop for futility or efficacy, drop an arm, select a dose, enrich a population, modify randomisation or re-estimate sample size. Thresholds and discretion boundaries should be clear enough to reproduce the action. Unplanned flexibility after results are seen threatens the reliability of the conclusion.

A trustworthy account of decision rules turn evidence into actions connects interim information, action and final inference. A rule may stop for futility or efficacy, drop an arm, select a dose, enrich a population, modify randomisation or re-estimate sample size. Thresholds and discretion boundaries should be clear enough to reproduce the action. Unplanned flexibility after results are seen threatens the reliability of the conclusion. Remember: Prospective adaptation is not permission to redesign a trial after comparative results are known.

A strong science note separates observation from interpretation. Challenge the timing, delayed outcomes, error control, estimation and operational firewall behind decision rules turn evidence into actions. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Repeated looks and multiple adaptation paths create more opportunities for an erroneous positive conclusion. The design uses suitable statistical methods to control the chance of false success across the complete strategy. Testing only the final-stage comparison as if no adaptation occurred can invalidate the type I error claim.

The beauty of multiplicity and error control remain central is disciplined learning. Repeated looks and multiple adaptation paths create more opportunities for an erroneous positive conclusion. The design uses suitable statistical methods to control the chance of false success across the complete strategy. Testing only the final-stage comparison as if no adaptation occurred can invalidate the type I error claim. Keep this boundary visible: Type I error control does not by itself guarantee unbiased estimation or operational integrity.

A strong science note separates observation from interpretation. Reconstruct multiplicity and error control remain central with protocol version, code, random seed, interim report and decision minutes. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A design can control type I error yet still yield biased point estimates or intervals after treatment selection or early stopping. The analysis should address the selection and stopping process and provide treatment-effect estimates with appropriate uncertainty, not only a pass-or-fail decision.

For estimation needs its own plan, every branch belongs to one prospective strategy. A design can control type I error yet still yield biased point estimates or intervals after treatment selection or early stopping. The analysis should address the selection and stopping process and provide treatment-effect estimates with appropriate uncertainty, not only a pass-or-fail decision. Preserve this guardrail: Simulation evidence is bounded by the scenarios, models and implementation code that were tested.

A strong science note separates observation from interpretation. Test estimation needs its own plan under null, target, heterogeneous, missing-data and time-trend scenarios. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Complex adaptive designs are often evaluated by simulation under null, target, heterogeneous, delayed-outcome and operational scenarios. Results include type I error, power, bias, interval coverage, sample size, duration and adaptation probabilities. Code, random seeds, assumptions and scenario grids should be reproducible.

A trustworthy account of simulation tests the whole strategy connects interim information, action and final inference. Complex adaptive designs are often evaluated by simulation under null, target, heterogeneous, delayed-outcome and operational scenarios. Results include type I error, power, bias, interval coverage, sample size, duration and adaptation probabilities. Code, random seeds, assumptions and scenario grids should be reproducible. Remember: E20 remained a draft in Step 3 on 27 October 2026.

A strong science note separates observation from interpretation. Trace simulation tests the whole strategy from interim data cut through statistic, threshold, authorised decision and final inference. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

If the primary endpoint arrives slowly, interim decisions may rely on incomplete follow-up or an earlier outcome. The relationship between early and final information, missingness and pipeline participants must be modelled. A fast but weak surrogate can create confident decisions that do not improve the final trial.

The beauty of delayed outcomes can weaken adaptation is disciplined learning. If the primary endpoint arrives slowly, interim decisions may rely on incomplete follow-up or an earlier outcome. The relationship between early and final information, missingness and pipeline participants must be modelled. A fast but weak surrogate can create confident decisions that do not improve the final trial. Keep this boundary visible: Prospective adaptation is not permission to redesign a trial after comparative results are known.

A strong science note separates observation from interpretation. Challenge the timing, delayed outcomes, error control, estimation and operational firewall behind delayed outcomes can weaken adaptation. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A trial may narrow recruitment to a biomarker-defined or more responsive subgroup. The design must protect inference for every population on which a claim may be made and avoid selecting a subgroup because of random high results. Prevalence, assay performance and overlap between populations affect operating characteristics.

For population enrichment changes interpretation, every branch belongs to one prospective strategy. A trial may narrow recruitment to a biomarker-defined or more responsive subgroup. The design must protect inference for every population on which a claim may be made and avoid selecting a subgroup because of random high results. Prevalence, assay performance and overlap between populations affect operating characteristics. Preserve this guardrail: Type I error control does not by itself guarantee unbiased estimation or operational integrity.

A strong science note separates observation from interpretation. Reconstruct population enrichment changes interpretation with protocol version, code, random seed, interim report and decision minutes. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Changing allocation toward better-performing arms may seem attractive, but time trends, delayed outcomes and operational behaviour can distort comparisons. The design should justify the rule, preserve enough concurrent control information and evaluate participant benefit rather than assuming unequal allocation is automatically more ethical.

A trustworthy account of response-adaptive randomisation can affect comparability connects interim information, action and final inference. Changing allocation toward better-performing arms may seem attractive, but time trends, delayed outcomes and operational behaviour can distort comparisons. The design should justify the rule, preserve enough concurrent control information and evaluate participant benefit rather than assuming unequal allocation is automatically more ethical. Remember: Simulation evidence is bounded by the scenarios, models and implementation code that were tested.

A strong science note separates observation from interpretation. Test response-adaptive randomisation can affect comparability under null, target, heterogeneous, missing-data and time-trend scenarios. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Knowledge of comparative interim results can change recruitment, assessment, adherence and care. Access is restricted to authorised independent groups, and communications reveal only what is needed to implement the decision. Roles, data transfers, meetings and unblinding events should be documented.

The beauty of information firewalls protect trial conduct is disciplined learning. Knowledge of comparative interim results can change recruitment, assessment, adherence and care. Access is restricted to authorised independent groups, and communications reveal only what is needed to implement the decision. Roles, data transfers, meetings and unblinding events should be documented. Keep this boundary visible: E20 remained a draft in Step 3 on 27 October 2026.

A strong science note separates observation from interpretation. Trace information firewalls protect trial conduct from interim data cut through statistic, threshold, authorised decision and final inference. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Interim datasets may be less cleaned or mature than final data. The plan identifies critical fields, query cut-offs, endpoint adjudication, missing-data handling and snapshot procedures. An automated dashboard is useful only if its inputs, algorithms and version are validated for the decision being made.

For data quality must support rapid decisions, every branch belongs to one prospective strategy. Interim datasets may be less cleaned or mature than final data. The plan identifies critical fields, query cut-offs, endpoint adjudication, missing-data handling and snapshot procedures. An automated dashboard is useful only if its inputs, algorithms and version are validated for the decision being made. Preserve this guardrail: Prospective adaptation is not permission to redesign a trial after comparative results are known.

A strong science note separates observation from interpretation. Challenge the timing, delayed outcomes, error control, estimation and operational firewall behind data quality must support rapid decisions. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Rules can pause enrolment, alter dose or stop treatment in response to safety information, but rare and delayed harms may not fit simple thresholds. Aggregate unblinded review, external information and benefit-risk judgement can be necessary. The procedure should distinguish urgent protection from efficacy-driven adaptation.

A trustworthy account of safety adaptation needs clinical judgement connects interim information, action and final inference. Rules can pause enrolment, alter dose or stop treatment in response to safety information, but rare and delayed harms may not fit simple thresholds. Aggregate unblinded review, external information and benefit-risk judgement can be necessary. The procedure should distinguish urgent protection from efficacy-driven adaptation. Remember: Type I error control does not by itself guarantee unbiased estimation or operational integrity.

A strong science note separates observation from interpretation. Reconstruct safety adaptation needs clinical judgement with protocol version, code, random seed, interim report and decision minutes. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

A single adaptive confirmatory trial sits beside earlier dose, biomarker, safety and mechanism evidence and possibly another confirmatory study. The design should not compress every learning objective into one trial when doing so makes estimation, integrity or replication too fragile.

The beauty of the development programme supplies context is disciplined learning. A single adaptive confirmatory trial sits beside earlier dose, biomarker, safety and mechanism evidence and possibly another confirmatory study. The design should not compress every learning objective into one trial when doing so makes estimation, integrity or replication too fragile. Keep this boundary visible: Simulation evidence is bounded by the scenarios, models and implementation code that were tested.

A strong science note separates observation from interpretation. Test the development programme supplies context under null, target, heterogeneous, missing-data and time-trend scenarios. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

The record includes draft and regional status, estimand, adaptation rationale, interim timing and data cut, decision rules, statistical methods, multiplicity control, estimation approach, simulation code and scenarios, governance and firewalls, actions taken, deviations, final analysis, sensitivity checks and development-programme context.

For a complete e20 receipt reconstructs every branch, every branch belongs to one prospective strategy. The record includes draft and regional status, estimand, adaptation rationale, interim timing and data cut, decision rules, statistical methods, multiplicity control, estimation approach, simulation code and scenarios, governance and firewalls, actions taken, deviations, final analysis, sensitivity checks and development-programme context. Preserve this guardrail: E20 remained a draft in Step 3 on 27 October 2026.

A strong science note separates observation from interpretation. Trace a complete e20 receipt reconstructs every branch from interim data cut through statistic, threshold, authorised decision and final inference. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter