When AI Becomes a Clinical Research Measure

Introduction

For decades, clinical research measures came from people, laboratories, and medical devices. Increasingly, they are coming from algorithms.
An AI system may estimate disease progression from a medical image, generate a mobility score from wearable data, or identify patterns that suggest a safety event. When those outputs influence participant assessment, endpoint determination, or study conclusions, AI stops being a research tool and becomes part of the evidence itself.

As AI becomes more common in digital health technologies, Software as a Medical Device (SaMD), remote monitoring systems, and advanced data-analysis platforms, sponsors, investigators, and institutional review boards are increasingly being asked to evaluate AI-generated measures. The U.S. Food and Drug Administration has likewise emphasized evaluating AI models based on their specific context of use and their role in supporting scientific and regulatory decision-making.

That distinction matters because an AI-derived measure must be credible enough for the decisions it informs. Researchers need to understand what the system measures, how consistently it performs, which participants it measures accurately, and how changes to the technology might affect study results.

The Central Principle

AI becomes a clinical research measure when its output contributes to participant assessment, eligibility determination, safety evaluation, endpoint determination, or study conclusions.

If the output can influence how a participant is evaluated or how a study result is interpreted, evaluate it as part of the evidence-generation process.

When Does AI Become a Clinical Research Measure?

A clinical research measure is information used to describe or evaluate a participant’s health, behavior, condition, response, or outcome. AI becomes a clinical research measure when its output contributes to one or more of the following:

• Participant assessment
• Eligibility determination
• Safety evaluation
• Treatment or study-management decisions
• Endpoint determination
• Study interpretation

A useful threshold question is:

If the AI output changed, could the interpretation of the participant’s condition or the study’s results also change?

If yes, evaluate the output as part of the study’s measurement and evidence-generation process.

From Observed Data to Algorithmic Measures

Historically, researchers generated many measures through direct observation, laboratory analysis, validated instruments, or established calculations. AI introduces an additional layer. Rather than measuring health directly, the system often interprets data collected elsewhere.

For example, a wearable device may collect raw movement data, but the study does not use those data directly. Instead, an algorithm converts the signals into a mobility score that investigators use.

The research team is therefore relying on more than the device itself. It is relying on the sensor, the data pipeline, the algorithm, and the evidence supporting the resulting measure.

This is why digital medicine frameworks increasingly separate verification, analytical validation, and clinical validation when evaluating whether a digital measure is fit for purpose. The Digital Medicine Society developed the V3 framework to evaluate digital clinical measures and determine whether they generate evidence suitable for their intended use.

AI-Derived Endpoints, Digital Biomarkers, and Digital Clinical Measures

AI is increasingly used to create digital clinical measures that may ultimately contribute to clinical trial endpoints.

Examples include:

• Gait speed derived from wearable sensors
• Cognitive assessments derived from speech analysis
• Disease progression estimates derived from imaging
• Risk scores derived from multiple health data sources
• Physiological measurements generated through remote monitoring technologies

A digital health technology collects information. An algorithm interprets that information. A clinical study may then use the resulting measure as part of an endpoint.

Each component must be evaluated.

For example:

• Does the sensor collect accurate data?
• Does the algorithm interpret the data correctly?
• Does the resulting measure meaningfully reflect the concept being studied?

The Digital Medicine Society describes analytical validation as determining whether an algorithm produces a correct, accurate, and reliable measure of the physiological or behavioral phenomenon of interest.

Consistency alone is not enough. A measure must also be clinically meaningful and appropriate for its intended context.

Defining Context of Use and Fitness for Purpose

The role an AI model performs determines how researchers should evaluate it. A system that helps prioritize records for review carries different implications than a system that determines participant eligibility or contributes to a primary endpoint.

The U.S. Food and Drug Administration’s January 2025 draft guidance on AI emphasizes evaluating model credibility within a defined context of use.

Research teams should understand:

• What question the model answers
• What data it receives
• What output it generates
• Who reviews the output
• Which decisions the output influences
• What could happen if it is wrong
• Whether it contributes to an endpoint or regulatory submission

Likewise, a strong accuracy rate does not automatically establish that a model is appropriate for a clinical investigation.

Researchers should evaluate:

• Whether validation data reflect the study population
• How the model performs across participant groups
• How it handles poor-quality or incomplete data
• Whether performance remains stable across sites and environments
• How to detect performance problems during the study

The required level of evidence should match the importance of the output and the consequences of error.

The Underlying Data Remain Part of the Measure

AI does not compensate for poor-quality data. An AI-generated measure may be affected by sensor placement, image quality, missing observations, software configuration, inconsistent documentation, or differences in how sites collect information. The U.S. Food and Drug Administration’s guidance on digital health technologies highlights the importance of selecting, documenting, and controlling technologies used to collect research data remotely.

Research teams should be able to trace:

• Data collection
• Data transmission
• Data preprocessing
• Algorithm processing
• Final output generation

Documentation should support reproducibility, quality review, and investigation of unexpected results.

Scenario: A Mid-Study Software Update

A trial uses wearable devices to generate a mobility score as a secondary endpoint. Halfway through the study, the vendor updates the scoring algorithm. The raw sensor data have not changed, but the interpretation of those data may have. Researchers must determine whether pre-update and post-update scores remain comparable and whether the update affects endpoint analyses.

A routine software update can become a measurement issue.

Bias Can Become Measurement Error

Bias in AI is not only an ethical concern. It can also become a measurement problem. Consider a study that uses speech analysis to estimate cognitive change. If the model performs differently across accents, languages, demographics, or recording conditions, the measure may reflect training-data characteristics as much as cognitive function.

Research teams should evaluate:

• Representation in development data
• Subgroup performance
• Potential proxy variables
• Historical biases in source data
• Differences in error rates across populations

These questions are not only technical and regulatory. They also intersect with longstanding principles of research ethics, including fairness, respect for persons, and participant protections.

Responsible AI principles such as fairness, accountability, transparency, privacy, reliability, and human agency are directly relevant to measurement quality and are explored in CITI Program’s Essentials of Responsible AI course.

Scenario: A Measure That Works Better for Some Participants

A study adopts a camera-based system to estimate a physiological measurement. Overall model performance appears strong, but performance varies among participant groups. If the study does not examine these differences, it may unintentionally introduce systematic measurement error.

The question is not simply whether the model works. It is whether it works consistently for the participants the study intends to enroll.

Transparency, Oversight, and Change Control

Research teams do not necessarily need access to proprietary source code. Still, they do need enough information to understand how an AI-generated measure is produced and how to use it.

Appropriate information may include:

• Intended purpose
• Input data sources
• Validation results
• Known limitations
• Failure conditions
• Human review requirements
• Effects of software or model updates

Human oversight must also be meaningful. Investigators should understand the output, recognize possible limitations, and know when independent review or escalation is appropriate.

Organizations should establish:

• Who reviews outputs
• What training or expertise is required
• How disagreements are documented
• How updates are evaluated
• How performance is monitored
• When the system should be suspended or revalidated

These issues are especially important when AI functions as Software as a Medical Device or clinical decision support. CITI Program’s Essentials of Software as a Medical Device & Clinical Decision Support Systems course addresses many of these regulatory and operational considerations.

Participant Protection Extends Beyond Technical Performance

A technically accurate model can still create ethical and operational concerns. AI-enabled research frequently involves large amounts of personal information, including physiological, behavioral, environmental, and location-related data. Participants may not realize the extent of the information inferred from those data.

Researchers should consider whether participants understand:

• What information is being collected
• How AI analyzes the data
• What the system is intended to infer
• Who receives the resulting outputs
• How long data are retained
• Whether the information may be reused

Researchers should also evaluate usability and accessibility. If a technology consistently places a greater burden on certain groups, it can reduce data quality and limit participant representation.

The Digital Medicine Society’s Applied Digital Health Ethics course addresses ethical risk identification, AI and machine learning harms, and equity considerations in digital health and research.

Seven Questions to Ask When Evaluating an AI-Generated Measure

Before relying on an AI-generated value as a clinical research measure, research teams should consider seven key questions.

• Purpose – What exactly is the model intended to measure, and how will the output be used?
• Evidence – Has the model been evaluated for the intended population, setting, and task?
• Data – Can the path from source data to final output be traced and explained?
• Performance – Are reliability, error rates, subgroup performance, and failure conditions understood?
• Oversight – Who reviews outputs, investigates problems, and assumes accountability?
• Change Control – How are updates documented, evaluated, and monitored?
• Participant Protection – How are privacy, ethics, accessibility, equity, and informed consent addressed?

The review should incorporate clinical, technical, statistical, ethical, regulatory, and data-management expertise.

For organizations working with sensor-derived measures, Building Fit-for-Purpose Sensor-based Digital Health Technologies: A Crash Course and The Playbook: Successfully Developing and Deploying Digital Clinical Measures offer additional resources on validation, deployment, and digital measurement strategy.

Training Supports Better Evaluation

Training cannot establish that a specific AI model is valid or compliant. It can help stakeholders ask better questions and identify when they need additional expertise.

As AI-generated measures become more common, research teams benefit from a shared understanding of:

• Context of use
• Digital clinical measures
• Validation concepts
• Data quality
• Bias and fairness
• Human oversight
• Software as a Medical Device
• Participant protections

Relevant offerings from CITI Program include Essentials of Responsible AI, Big Data and Data Science Research Ethics, and Essentials of Software as a Medical Device & Clinical Decision Support Systems, along with digital-health and digital-measurement content from the Digital Medicine Society series.

Evaluate the Measure, Not the AI Label

Terms such as “AI assistant,” “decision-support tool,” “digital biomarker,” and “predictive model” do not determine how consequential a technology is. Its role in the study does.

When an AI output contributes to participant assessment, safety determination, endpoint evaluation, or study interpretation, it becomes part of the evidence-generation process. Researchers should evaluate whether the measure is meaningful, evidence-supported, derived from trustworthy data, subject to appropriate oversight, and suitable for the decisions it informs.

AI may enable new ways to evaluate health and generate evidence, but the output is not credible simply because it was produced by an advanced algorithm. It is whether the study can explain, support, and responsibly rely on what the AI is measuring.