Artificial Intelligence Transforms Clinical Trials, but Challenges Remain Enormous

Artificial intelligence is gradually establishing itself in the field of clinical trials, promising to accelerate medical research, reduce costs, and improve the effectiveness of treatments. However, its integration raises complex questions that go beyond mere technological promises.

In protocol design, artificial intelligence now makes it possible to automatically generate complex documents from simple summaries provided by researchers. This significantly reduces the time required to draft detailed protocols, a task that was once long and tedious. However, this speed comes with risks, particularly the potential disclosure of sensitive information about drugs or innovative designs when this data is entered into external systems. Predictive models, on the other hand, analyze historical databases to anticipate key operational outcomes, such as the ability to meet recruitment goals or identify optimal inclusion criteria. These tools help avoid costly failures, but their effectiveness heavily depends on the quality and representativeness of the data used.

Patient recruitment, often a source of delays in trials, also benefits from the contributions of artificial intelligence. Advanced systems analyze millions of medical documents in record time to identify eligible candidates. For example, one platform processed over 83 million documents in less than two weeks, generating billions of standardized medical terms. These technologies can cut manual screening time in half, freeing up time for medical teams. They also reveal that many eligibility criteria unnecessarily exclude patients without affecting therapeutic outcomes. By relaxing these rules, trials could include twice as many participants while maintaining their safety.

However, the performance of these systems varies. Some achieve high recall rates, identifying the majority of eligible patients, but with sometimes low precision, meaning they also include ineligible candidates. Conversely, a highly precise system reduces the workload for human reviewers but risks excluding valid patients. Recent advancements, such as models combining data retrieval and generation, show promising results with accuracy rates exceeding 97%, even surpassing human coordinators for certain tasks.

Data analysis is another area where artificial intelligence is making strides. It enables the processing of complex datasets, such as those from connected devices or medical imaging, providing unprecedented richness of information. Yet, a debate persists between traditional statistical methods and modern models. Classical approaches, like logistic regression, offer unmatched transparency, allowing a clear understanding of the influence of each variable. Artificial intelligence models, while often more accurate, operate as black boxes, making their internal reasoning difficult to decipher. This opacity poses a major challenge for clinical adoption, as physicians hesitate to follow recommendations they do not understand, especially in critical situations.

Algorithmic biases represent another significant obstacle. Systems can perpetuate, or even amplify, existing inequalities if the data used for their training reflects historical disparities. For example, an algorithm widely used in the United States to assess intensive care needs favored white patients over Black patients because it used past healthcare costs as an indicator of future needs, a criterion biased by structural inequalities. Similarly, cardiovascular diagnostic tools trained primarily on male data show reduced accuracy for women, whose symptoms often differ. These biases highlight the importance of diversifying development teams and datasets to ensure fair outcomes.

Model transparency is also a crucial issue. Clinicians and patients must be able to understand how a decision was made, especially when it directly impacts lives. Post-hoc explanation methods, such as LIME or SHAP, attempt to make models more understandable by identifying the most influential factors for a given prediction. However, these explanations remain approximations and can themselves be misleading. Some models, designed to be inherently interpretable, such as decision trees, offer an alternative, although their accuracy may be lower than that of complex models.

The challenges are not just technical. The integration of artificial intelligence into clinical trials raises unresolved ethical and legal questions. Who is responsible in the event of a medical error involving an artificial intelligence system? The developer, the medical institution, the clinician, or the patient? The diffusion of this responsibility, often referred to as the “many hands problem,” complicates fault attribution and may leave patients without clear recourse. Some propose a collective responsibility model, where all involved parties share the burden, but this approach remains underdeveloped.

Data ownership and patient consent also pose problems. In a traditional trial, a patient consents to the use of their data for a specific and time-limited purpose. With artificial intelligence, data is often reused indefinitely to train and improve models, raising questions about its ownership and control. Data sovereignty models, where patients retain granular control over their information, are emerging as a possible solution, but their large-scale implementation remains complex.

Decentralized systems, such as federated learning and blockchain, offer avenues to overcome obstacles related to privacy and data access. Federated learning allows models to be trained on data distributed across multiple institutions without ever centralizing sensitive information. However, this approach faces practical challenges, such as data heterogeneity across sites, which can degrade the performance of the global model. Blockchain, for its part, could ensure the integrity and traceability of data, but its adoption is hindered by issues of scalability, energy consumption, and compliance with data protection regulations, such as the right to be forgotten.

The costs and efficiency gains promoted by artificial intelligence are often highlighted. Studies show that online recruitment reduces costs per patient from $199 to $72 and significantly accelerates the process. Other reports indicate a 10 to 15% reduction in enrollment times in pilot trials. However, these figures often come from industry-funded studies or internal reports, which may be biased or lack independent validation. Rigorous economic analyses remain rare, and the available results show significant methodological gaps.

Finally, the impact of artificial intelligence on clinicians’ workload and patient experience is ambivalent. While some tools automate tedious administrative tasks, freeing up time for direct care, others may instead increase the burden by requiring constant verification of generated results. Patients, on the other hand, oscillate between hope for more accurate diagnoses and personalized treatments, and fear of privacy loss or reduced human interaction. A recent survey reveals that 60% of Americans would be uncomfortable if their healthcare provider relied on artificial intelligence for medical decisions.

Documentation and Sources

Reference Document

DOI: https://doi.org/10.1007/s10791-026-10205-x

Title: The transformative and turbulent integration of artificial intelligence in clinical trials: a critical review

Journal: Discover Computing

Publisher: Springer Science and Business Media LLC

Authors: Ying Xuan Lim; Long Chiau Ming; Nancy Choon-Si Ng; Serena Leow; Rahul G. Ingle