Medicines don’t work the same for everyone, which means clinical trials shouldn’t look the same for everyone either.

Clinical trial diversity is essential for advancing health equity and understanding how medications work differently for different populations. A lack of diversity can limit the generalizability of trial results, slow enrollment, and worsen health disparities, which can have significant consequences.

Artificial intelligence (AI) is being used to improve clinical trial diversity in various ways, such as identifying eligible participants and supporting targeted recruitment, as well as helping to reduce bias when recruiting participants. AI can analyze patient data to match eligible participants with clinical trial criteria, improving efficiency in recruitment by significantly decreasing the time of an otherwise lengthy process.

However, existing data sets are used to train the majority of these AI models, and many of those data sets are incomplete or contain historical biases. This can lead algorithms to unintentionally reinforce existing disparities, demonstrating the importance of validation and bias assessments.

AI Enhancing DEI in Clinical Trials

Randomized clinical trials were long considered the gold standard, but they’re only one type of study design. Many other approaches such as observational studies, pragmatic trials, adaptive designs, platform trials, and decentralized models, introduce different strengths and limitations for reaching diverse populations. When these designs rely on biased historical data or narrow site networks, they can still fail to represent the communities the treatment is intended to serve. AI can support more inclusive study designs, monitor diversity in real time, and help reach underrepresented populations. When applied thoughtfully, it can also strengthen the relevance of clinical trial results across diverse groups.

A study published by the National Library of Medicine found that AI-powered patient recruitment tools have improved enrollment rates by 65%, and are able to detect hidden biases, forecast recruitment shortfalls, and recommend corrective measures in real time. AI-powered tools can be used to expand outreach by identifying underserved populations and customizing communication strategies in order to reduce cultural and socioeconomic barriers to participation.

AI can increase clinical trial diversity in a number of ways, most notably:

• Patent Identification: AI can use health records and health data to identify participants who are eligible for the study, particularly groups of people who may have been overlooked. An AI-powered screening tool was found to screen out 72% of ineligible patients from the initial pool, which used to be a very time-consuming process.
• Targeted Recruitment: AI can be used to identify underrepresented communities within a study, so researchers can focus on outreach where it is necessary. A study analyzing the effects of AI on recruitment found that AI increases efficiency, cost savings, accuracy, patient satisfaction, and creates user-friendly interfaces.
• Reducing Bias in Recruitment: By flagging patterns of demographic groups that are being underrepresented, AI can reduce bias in recruitment. Algorithms are able to identify populations that are underrepresented in real-world datasets and utilize that information during participant outreach.
• Participant Matching: AI can analyze participants’ medical history to match them to the appropriate clinical trial, without needing a referral. AI-powered patient matching was able to identify eligible patients with 78% accuracy, compared to 45% accuracy with traditional screening methods.

Challenges of AI on Clinical Trial Diversity

AI can enhance clinical trial diversity but it also presents significant challenges. Bias in model training can shape =eligibility recommendations, virtual trial approaches may miss key populations, and unvalidated outputs can erode trust.

Much of this stems from the data AI learns from. Clinical research datasets and EHRs already contain historical bias; underrepresentation of certain groups, site patterns that favor specific populations, and eligibility criteria built without diverse participants in mind. When AI models are trained on these non-representative datasets, they inherit and amplify those gaps, producing predictions that are less accurate for some patient groups.

Bias also emerges when proxy variables stand in for unmeasured factors. TThese shortcuts can distort outcomes, as seen in a widely used commercial algorithm that used healthcare costs as a proxy for illness severity, resulting in Black patients being systematically under-referred to specialized care.

Best Practices (Strategies to Address Algorithmic Bias in AI Models)

Given the evident challenges of using AI, validation and bias assessments are crucial. Strategies to address algorithmic bias in AI models include training models using diverse datasets, ensuring the accuracy of labeled data, establishing standardized reporting for AI training data and limitations, and conducting a thorough evaluation of model performance.

These practices, when used together, help to ensure fairness, transparency, and reliability of AI models, reducing the potential for biased outputs and supporting equitable participant identification and decision-making in clinical trials.

How Clarkston Can Help

• AI Readiness Assessment: We evaluate your current state and build a clear roadmap for what needs to be in place before AI-driven recruitment is even
• Vendor Evaluation and Selection: We help you ask the right diversity and bias questions, so you can choose partners who can truly support equitable recruitment before = any contract is signed.
• Data and Strategy Foundation: We establish the right data structures, processes, and governance so that when AI arrives it is built on fair, high-quality, representative data from day one.

Contributions from Kate Sinha