03 · What You Need to Know
How AI Should Be Used in Research Stopping and Safety Decisions
What Does It Mean to Stop a Research Study?
Stopping a research study can refer to several different actions. These include temporarily suspending recruitment, pausing an intervention, discontinuing a particular study arm, stopping data collection, or permanently terminating the entire study.
These actions have different consequences and may be governed by different procedures.
For example, a research team might temporarily pause recruitment while investigating a safety concern without ending follow-up for participants already enrolled.
Alternatively, an independent monitoring committee might recommend stopping a randomized trial early because a prespecified statistical boundary has been crossed.
Researchers should distinguish suspension from permanent termination and identify exactly which activity an AI-generated recommendation concerns.
A recommendation to stop administering an intervention is not necessarily a recommendation to discontinue all participant monitoring or data collection.
Why Might a Study Be Stopped Early?
Research studies may be stopped or modified for several reasons, including participant safety, compelling evidence of benefit, futility, operational infeasibility, or ethical and regulatory concerns.
| Reason |
Central concern |
Relevant decision process |
| Safety |
Emerging evidence suggests unacceptable participant risk |
Safety assessment, protocol requirements, and authorized protective action |
| Efficacy |
Interim evidence may justify concluding that an intervention is beneficial |
Prespecified monitoring boundaries and evaluation of the total evidence |
| Futility |
Continuing may have a low prospect of answering the intended question or meeting a defined objective |
Appropriate conditional power, predictive probability, or other planned criteria |
| Feasibility |
Recruitment, resources, or operational problems threaten completion |
Investigator, sponsor, and institutional assessment |
| Ethical or regulatory concern |
The study may no longer satisfy applicable requirements |
Relevant ethics, regulatory, and institutional procedures |
These categories may overlap. A study may experience safety concerns and recruitment difficulties simultaneously.
AI should not treat all stopping reasons as equivalent or assume that a single statistical indicator determines the appropriate action.
Safety Monitoring Requires More Than Detecting Unusual Data
AI systems may help identify patterns in adverse event reports, clinical measurements, participant communications, or other safety-related information.
For example, a system might flag an unexpected increase in a particular adverse event or identify records requiring urgent professional review.
However, an unusual pattern does not necessarily establish that the intervention caused harm.
Researchers may need to examine event severity, timing, alternative explanations, exposure, background rates, and the reliability of the underlying data.
False-positive alerts may lead to unnecessary interruptions, while false negatives may delay protective action.
The separate question of using AI to flag participant safety risks requires appropriate validation and escalation procedures.
AI-generated alerts should be treated as signals for investigation unless a validated and authorized procedure establishes a more specific role.
Interim Analyses Require Prespecified Statistical Planning
An interim analysis examines accumulating study data before the planned final analysis.
In randomized trials, repeated examination of treatment differences can affect the probability of false-positive conclusions if conventional testing thresholds are applied without appropriate adjustment.
Group sequential designs address this concern by specifying monitoring procedures and statistical boundaries across planned analyses.
For example, an O'Brien-Fleming-type design generally uses relatively stringent early efficacy boundaries that become less stringent at later planned analyses.
Other approaches may use different monitoring functions, Bayesian decision rules, or adaptive procedures.
The appropriate design depends on the trial's objectives, assumptions, and regulatory context.
AI may help implement prespecified calculations, but it should not invent interim stopping thresholds after examining accumulating results.
Stopping for Efficacy Is Not Simply Finding P Below 0.05
A study may produce a statistically significant interim result without satisfying the planned stopping boundary.
Repeated testing can increase the risk of false-positive conclusions when multiplicity is not appropriately controlled.
Furthermore, early estimates of treatment benefit may be unstable or exaggerated, particularly when a study stops following an unusually favorable interim result.
Researchers should evaluate the relevant monitoring boundary, effect estimate, uncertainty, safety information, and the totality of available evidence.
AI should not recommend stopping solely because an unadjusted interim p-value crosses 0.05.
The distinction between statistical significance and scientific interpretation remains essential.
Stopping for Futility Requires Careful Interpretation
Futility monitoring evaluates whether continuing a study is likely to achieve a specified scientific or statistical objective.
Conditional power estimates the probability of meeting a future testing criterion under specified assumptions about the remaining data. Bayesian predictive probability provides a related but distinct approach under an explicit probabilistic model.
Neither quantity is a universal measure of whether a study is worth continuing.
A low conditional power estimate may depend strongly on assumptions about the treatment effect and future observations.
Researchers should also consider safety, feasibility, scientific relevance, and whether the planned futility rule is binding or nonbinding.
AI-generated recommendations should not conceal these assumptions behind a simple instruction to stop.
Data Monitoring Committees Provide Independent Oversight
Some clinical trials use an independent data monitoring committee, also called a data and safety monitoring board, to examine accumulating evidence and recommend whether the study should continue, change, or stop.
The committee's responsibilities, access to unblinded information, and decision procedures are typically defined in a charter or equivalent document.
Not every research study requires such a committee. Its use depends on the nature of the intervention, participant risks, trial design, sponsor requirements, and applicable regulations.
Where a committee is established, AI may support preparation of analyses or identification of issues requiring examination.
It should not replace the committee's independent judgment or the authority of the individuals and organizations responsible for implementing decisions.
The U.S. Food and Drug Administration provides guidance on establishing and operating clinical trial data monitoring committees.
Human Authority Must Be Defined Before AI Is Introduced
Researchers should establish who can recommend, authorize, and implement study suspension or termination.
Depending on the research setting, relevant parties may include the principal investigator, sponsor, independent monitoring committee, ethics committee, institutional officials, or regulatory authority.
These roles are not interchangeable.
For example, a monitoring committee may recommend stopping a trial, while the sponsor or other authorized party is responsible for the formal decision and implementation under the applicable arrangements.
Investigators may also have immediate responsibilities to protect participants when urgent safety concerns arise.
AI should not be assigned decision-making authority that conflicts with the approved protocol or applicable governance requirements.
Urgent Safety Action May Be Necessary Before Full Confirmation
Researchers should not interpret the requirement for human oversight as a reason to delay urgent participant protection.
Where credible evidence suggests an immediate serious risk, an approved safety procedure may require prompt action before the concern has been fully investigated.
For example, an investigator may need to suspend a potentially hazardous procedure while obtaining appropriate clinical or institutional assessment.
The response should follow the study's emergency procedures and applicable requirements.
An AI alert may contribute to identifying the concern, but the system should not independently diagnose the cause or determine the final disposition of the study.
Researchers must balance the risks of unnecessary interruption against the potentially greater consequences of delayed protective action.
AI False Positives and False Negatives Can Affect Stopping Decisions
An AI system that incorrectly identifies a safety signal may contribute to unnecessary suspension, participant anxiety, resource loss, or premature termination.
Conversely, a system that fails to identify a genuine signal may delay appropriate intervention.
The consequences of these errors may be asymmetric.
Researchers should examine relevant performance measures, including sensitivity, specificity, calibration where appropriate, and the system's behavior in the intended population.
However, predictive performance alone does not establish whether an AI-supported stopping procedure is ethically acceptable.
The broader principles governing AI errors that could materially harm participants should guide the assessment.
Study Stopping Decisions May Affect Scientific Validity
Stopping a study early can affect the precision of estimates, the interpretation of treatment effects, and the availability of information about longer-term outcomes.
In some circumstances, early termination may leave important safety or effectiveness questions unresolved.
Continuing a study despite compelling evidence of harm may also be scientifically and ethically indefensible.
Researchers should therefore evaluate the consequences of both stopping and continuing.
AI recommendations should not focus exclusively on one statistical indicator while ignoring participant welfare or the study's broader objectives.
AI Model Changes Can Undermine Monitoring Consistency
If AI is used to identify safety signals or classify adverse events, changes to the model may alter which cases are flagged.
A system update could affect sensitivity, specificity, or the interpretation of previously established alert thresholds.
Researchers should document the model and relevant settings, evaluate material changes, and determine whether revalidation is necessary.
For high-stakes monitoring, a major unassessed change may justify temporarily suspending the AI component while maintaining necessary safety oversight through established procedures.
The continuity of participant protection should not depend on an external AI service remaining unchanged.
Stopping Procedures Must Be Auditable
A defensible stopping decision should be supported by records identifying the triggering evidence, relevant analyses, applicable protocol criteria, and authorized decision-makers.
Where AI contributes, researchers should document the system's role, the information it processed, its output, and how that output was independently assessed.
Records should distinguish an AI-generated alert from a confirmed safety finding and from the formal decision to suspend or terminate the study.
Researchers should also document notifications, participant protections, and any required follow-up activities.
Watch Out
Do not allow a general-purpose AI system to invent stopping boundaries, interpret interim p-values without the approved monitoring plan, or independently terminate a study. Stopping decisions must follow the applicable scientific, ethical, and governance procedures.
