AI in Clinical Trials: Disentangling the “Noise” of Patient Behavior

Patient-related variability in clinical trials is driven by a combination of biological, environmental, and behavioral factors. While biological variability has long been studied and increasingly controlled through stratification and biomarkers, behavioral variability remains significantly under-characterized.

This type of variability includes factors such as patient expectations, motivation, perception of symptoms, adherence, and interaction with site staff. These elements can materially influence outcomes, particularly in trials relying on subjective or patient-reported endpoints. For example, two patients with identical biological responses to a treatment may report very different outcomes due to differences in optimism, mood, or expectations.

Historically, this layer of variability has been difficult to quantify for two reasons. First, behavioral traits are inherently multidimensional and dynamic, so they are not directly observable in the same way as biological markers. They often evolve over the course of a study. Second, traditional clinical trial infrastructures were not designed to capture behavioral data. As a result, these signals have remained embedded in “noise,” contributing to placebo response, dropout, and outcome variability without being explicitly modeled.

Only recently, with the combination of validated behavioral science frameworks and structured data collection, has it become possible to translate these previously latent factors into measurable variables that can be integrated into trial design and analysis.

AI can play a meaningful role in clinical development, but mainly as an enabler, dependent on the quality and relevance of the underlying data.

“When applied to robust, well-characterized datasets, AI enables the transformation of complex inputs such as behavioral data into actionable insights.” — Dr. Dominique Demolle.

In the context of clinical trials, this creates three main opportunities. First, in trial design, predictive models can help identify and quantify sources of variability that would otherwise remain unmodeled, allowing for covariate adjustments. This directly improves statistical power without necessarily increasing sample size.

Second, during trial execution, predicting patient behavior supports proactive risk management by identifying patients, sites, or regions at higher risk of disengagement, non-adherence, or dropout. Rather than reacting to issues once they materialize, study teams can intervene earlier and more selectively.

Third, in data interpretation, those predictive models can help disentangle signal from noise by modeling factors such as placebo response. This strengthens the reliability of trial outcomes and reduces the risk of inconclusive results.

Importantly, the true value of AI in this context is not automation alone, but the ability to make previously unmeasured drivers of variability visible, quantifiable, and ultimately actionable within existing clinical development workflows.

One key limitation is the lack of a clearly defined use case. AI models are only meaningful when they are designed to address a specific, well-formulated problem. Without this clarity, there is a risk of generating outputs that are not actionable nor aligned with study objectives.

Second, measurability and validation are critical. AI-driven approaches must demonstrate quantifiable impact on predefined endpoints, and repeatability, whether statistical, operational, or clinical. This includes ensuring that models are trained on relevant data and validated within the appropriate context of use.

Third, scalability and operational integration remain significant constraints. For AI to deliver value, it must fit seamlessly into existing workflows, analysis plans, and regulatory frameworks. Solutions that require process disruption or cannot be operationalized at scale are unlikely to be adopted.

Finally, there are considerations related to data quality and bias. AI models reflect the data they are trained on, so incomplete, unstructured, or biased datasets can limit reliability and generalizability.

“In practice, adoption may remain a barrier. AI-driven approaches typically require changes to established practices, and change is often perceived as carrying additional risk.” — Dr. Dominique Demolle.

It is essential to accurately assess the risk-benefit balance to reassure clinical teams and end users about the relevance and added value of this shift.

Regulators are increasingly supportive of AI-driven approaches in clinical development, provided they are applied within a transparent and well-defined framework.

A key reference point is the draft Food and Drug Administration guidance on the use of AI to support regulatory decision-making. The guidance outlines how AI-generated evidence can support regulatory decisions related to safety, efficacy, and quality.

Importantly, it introduces a risk-based credibility assessment framework, which evaluates AI models based on their specific context of use. This means that the level of evidence required depends on how the model is being used. For example, whether it informs exploratory insights or directly supports regulatory claims.

“The emphasis from regulators is not on the technology itself, but on credibility, transparency, and validation. Models must be explainable, appropriately validated, and integrated into the overall evidentiary package in a way that can be reviewed and understood.” — Dr. Dominique Demolle.

Overall, the direction is clear: regulators are not only open to AI but are actively defining the frameworks that will enable its safe and credible adoption in clinical development.

Transparency is not optional; it is a prerequisite for adoption in clinical research. We operate in an industry that directly affects patients’ lives, which leaves very limited tolerance for black-box approaches.

There is a tension between interpretability and performance, but in this context, performance without explainability has limited value. Determining the right balance between the two is critical. Approaches that combine strong performance with transparent, clinically interpretable outputs are more likely to gain regulatory and operational acceptance.

Access to data is equally fundamental. All companies face the same challenge: achieving the right balance between data sharing and protecting their competitive advantage. Sponsors, contract research organizations, technology providers, and regulators all generate valuable data, yet access remains fragmented. Breaking down these silos is not just an efficiency gain; it is essential to unlocking AI’s full potential. Global data-sharing platforms are important steps in the right direction, but their impact remains constrained by limited access, driven by privacy requirements, IP, and competitive concerns.

Compared with other industries, the pace of AI adoption in pharmaceuticals has been more measured, largely due to regulatory and data constraints. However, the direction is unequivocal: this sector will increasingly converge toward broader AI integration.

Overcoming these challenges is not simply a matter of alignment; it is a prerequisite for trust. Aligning scientific rigor, regulatory expectations, and ecosystem-wide collaboration is what will determine whether AI-driven approaches can move from promising innovation to credible, scalable tools.