AI and innovation in clinical trials

Abstract

Clinical trials face persistent challenges in cost, enrollment, and generalizability. This perspective examines how artificial intelligence (AI), large language models (LLMs), adaptive trial designs, and digital twins (DTs) can modernize trial design and execution. We detail AI-driven eligibility optimization, reinforcement learning for real-time adaptation, and in silico DT modeling. Methodological, regulatory, and ethical hurdles are addressed, emphasizing the need for validated, scalable frameworks to enable responsible and widespread integration.

Introduction – Challenges in Traditional Clinical Trials

Clinical Trials (CTs) are the gold standard in evidence-based medicine, but their successful design and implementation face several challenges. Substantial operational costs associated with a well-executed CT are a barrier to entry for many therapies. Despite feasibility efforts, many trials struggle with accrual. Elaborate protocols and stringent eligibility criteria impede the efficiency of participant identification and recruitment. For trials that do complete enrollment, patient cohorts may not be representative of the broader real-world patient populations that may benefit from the trial treatments. Moreover, the public’s perception of CTs—complicated by the complex terminology and schedule of events—can lead to hesitancy in enrolling or maintaining adherence. Thus, there is a pressing need for innovative solutions. Advances in Artificial Intelligence (AI) algorithms, including machine learning (ML) and Large-Language Models (LLMs), together with high-performance computing, hold significant potential to disrupt the status quo.

AI and Eligibility Optimization

Carefully designed eligibility criteria minimize patient harm, enhance real-world applicability, and ensure meaningful results. However, many eligibility parameters are in place due to heightened risk aversion (e.g., minimizing the inability to complete or tolerate the study treatment) or are carried forward from past protocols (e.g., lab tests for eligibility that may not be relevant to the study treatment)¹. Restrictive eligibility criteria frequently exclude significant portions of the population who might otherwise benefit from the intervention. Liu et al. address this issue by applying their ML-based algorithm (Trial Pathfinder) on 10 completed Phase III advanced non-small-cell lung cancer (aNSCLC) trials—including FLAURA, LUX8, Checkmate017, Checkmate057, Checkmate078, Keynote010, Keynote189, Keynote407, BEYOND, and OAK—by systematically comparing the eligibility criteria of these trials to data from the Flatiron Health EHR database, a curated, cleaned, and harmonized clinical dataset, which includes 61,094 NSCLC patients². They demonstrate that exclusions based on specific laboratory values often have minimal impact on trial outcomes like hazard ratios of overall survival. By applying a data-driven approach to broaden these criteria, they show that the pool of eligible patients doubled on average without compromising patient safety. These findings align with broader efforts to modernize CT eligibility frameworks using real-world data and AI models that reflect population diversity and comorbidity burdens^3–6. Training ML models on historical trial data can highlight which criteria are truly necessary, thereby improving real-world relevance and expanding equitable access.

Adaptive Trial Designs Enhanced by AI

The rapid development of multiple therapeutic options for a given condition, particularly rare conditions for which a vast pool of participants may not be available, demands alternative trial designs. Here, a “fail fast” strategy becomes essential. Adaptive trial designs meet this need by allowing pre-planned, real-time modifications to trial protocols, ensuring that the most promising interventions are quickly identified and prioritized.

Adaptive trial designs are inherently valuable because they allow modifications during the trial, such as dosage adjustments, adding or removing treatment arms, and reallocating patients based on their responses⁷. These designs enable the parallel testing of multiple candidate therapies, thereby accelerating the identification of promising interventions. ML algorithms can be leveraged to complement adaptive designs by rapidly analyzing large datasets to estimate outcomes and inform real-time adjustments to trial parameters⁸. Specifically, algorithms based on reinforcement learning, decision trees, and neural networks are then employed to continuously update the trial protocol based on interim results (Fig. 1). To ensure the statistical validity of such changes, Bayesian frameworks and frequentist group-sequential methods are often embedded within these adaptive designs. AI-enhanced reinforcement learning models can be aligned with such statistical thresholds by incorporating posterior probability distributions into the learning loops, thus maintaining type I error control^9,10. Reinforcement learning optimizes decision-making by evaluating potential outcomes of different adjustments, decision trees help identify the most effective treatment pathways, and neural networks process complex patient data to forecast responses to interventions. The resulting adaptability leads to more efficient trials by focusing resources on promising therapies earlier in the process, enabling the early identification and discontinuation of less promising options¹¹.

Fig. 1. Machine learning–driven adaptive trial design framework.

This schematic illustrates how machine learning (ML) algorithms can be integrated into adaptive clinical trial workflows. Interim results from ongoing trials are analyzed via reinforcement learning (to optimize decisions), decision trees (to identify optimal pathways), and neural networks (to forecast responses). These outputs are combined to inform real-time protocol updates—enabling dosage adjustments, treatment arm modifications, and patient reallocation.

Digital Twins – Toward Predictive and Synthetic Trial Architectures

Complementing the dynamic nature of adaptive trials, digital twins (DTs) introduce the potential to refine clinical research by investigating the most promising treatments based on characteristics of individual patients – i.e., n of 1 CTs. By simulating patient-specific responses, DTs could enhance treatment precision while also optimizing resource allocation throughout the trial process. They also address the legitimate concern of participants not wanting to be randomized to the control/placebo arm¹². In the context of CTs, a DT is a dynamic virtual representation of an individual patient created through integration of real-world data streams and computational modeling, which encompasses and simulates a patient’s physiological characteristics, disease features, and potential responses to treatments. Using datasets encompassing clinical, genetic, and lifestyle information, DTs can predict how a patient might respond to a specific treatment, ultimately giving way to personalized treatment regimens^3,12,13. Additionally, DTs can integrate multi-omics data and real-time health monitoring, providing a continuous and dynamic representation of an individual’s health status^14,15. Beyond individual-level simulations, DTs can also be developed for patient subgroups or populations using data from real-world CTs and observational studies. These group-level DTs can simulate different trial designs, help identify likely sources of failure, and refine protocols before trials are launched—ultimately improving efficiency and reducing the risk of costly trial failures.

However, a key methodological challenge in applying DTs to CTs is the quantitative comparison between predicted and actual patient outcomes. Retrospective validation has emerged as a pragmatic first step, where DT predictions are evaluated against completed trial data to measure performance gaps and identify efficiency gains. For example, DT simulations applied to Alzheimer’s datasets have shown alignment with historical patient trajectories, helping to assess surrogate endpoints and trial enrichment strategies¹⁶. Real-time validation is another promising strategy, enabling the integration of continuously updated electronic health records (EHRs) or wearable sensor data to dynamically recalibrate DT predictions, a process conceptually aligned with adaptive trial design¹⁷. To further ensure robustness, DT models increasingly employ prognostic covariate adjustment frameworks such as PROCOVA-MMRM to reduce sampling bias and improve power in longitudinal trials^18–20. In oncology and chronic disease modeling, predictive DTs combining imaging and mechanistic modeling now include quantified uncertainty to support patient-specific decision-making^21,22. The comparison of DT-predicted versus observed trajectories is frequently performed using survival concordance indices, RMSE, or calibration curves, often supported by mixed-effects or Bayesian models that accommodate population heterogeneity^23,24.

However, despite advances in multi-modal fusion techniques, gaps in real-time data—caused by incomplete EHRs, missing wearable inputs, or asynchronous clinical reporting—remain a major limitation. Recent innovations such as TWIN-GPT have demonstrated the ability to impute missing values and synthesize patient trajectories from sparse datasets by learning cross-dataset representations, enabling more complete and personalized twin generation even when real-time data are limited²⁵. Generalizability is especially difficult when DTs trained on narrowly defined cohorts are applied to broader clinical settings where underlying biology, comorbidities, or care standards differ. Addressing these limitations requires not only high-performance computing but also transparent, disease-specific validation pipelines capable of benchmarking performance at scale²⁶.

A different application of DTs within CTs is to enable synthetic control arms, where simulated patient data can replace or augment actual controls^27,28. In-silico CTs utilizing simulated control and efficacy arms could ultimately optimize patient recruitment and drug protocols. This fosters data-driven decision making earlier in the process, reducing costly failures and allowing resource conservation for more successful and efficient trials^3,23,29. While digital twins simulate individualized responses and trial structures, recent advances in AI agents introduce the possibility of real-time coordination across the trial lifecycle—enabling AI systems to act as orchestrators rather than passive predictors.

AI Agents – Autonomous Coordination Across the Clinical Trial Lifecycle

AI agents extend beyond traditional ML and LLMs by enabling autonomous, goal-directed actions across the CT lifecycle. These systems have the potential to coordinate activities leveraging trial protocols and patient data, invoking tools dynamically, and interacting with multiple components of a CT pipeline—mimicking roles such as coordinators or analysts.

Recent frameworks underscore their promise and value. ClinicalAgent, a multi-agent LLM system, improved trial outcome prediction by 0.33 AUC over baseline prompting methods by integrating real-world data and protocol reasoning³⁰. MAKAR achieved 100% accuracy in controlled patient-trial matching tasks by using role-specific agents to navigate complex eligibility criteria³¹. In operational contexts, oncology-focused GPT-4 agents with multimodal tool access reached 87% accuracy in diagnostic and enrollment decisions—nearly tripling the performance of standalone LLMs (30%)³².

The strength and risks of AI agents lies in their autonomy and interactivity. Unlike static models, they can monitor trial progress, adapt eligibility logic in real time, and trigger statistical tools or database queries. As such, they offer a promising layer of infrastructure to streamline operations, accelerate adaptive decisions, and reduce trial delays—complementing tools like DTs and reinforcement learning frameworks. At the same time, AI agents increase the complexity of model lifecycle management, necessitating safeguards to identify and mitigate unintended risks.

Infrastructure, Scalability, and Simulation Costs

Recent advances in high-performance computing and cloud-based technologies now enable these complex simulations to be performed at scale by providing on-demand computational storage, and secure data-sharing capabilities, facilitating the exploration of numerous treatment scenarios concurrently³³. Platforms such as AWS, Google Cloud, and Microsoft Azure offer environments that allow researchers to run complex in-silico trials without requiring extensive in-house infrastructure; however, these cloud-based services may not necessarily result in cost savings and can sometimes be significantly more expensive compared to on-premises computing. Despite potential cost implications, this simulation-driven strategy remains valuable as it can substantially reduce clinical risks and accelerate the development of more tailored and effective interventions.

Model Quality, Data Bias, and Interpretability

Despite these benefits, AI and DT implementation face significant challenges (Table 1). The accuracy and performance of AI models and DTs depend on the quality and completeness of the data^13,34,35. Incomplete, biased, or non-representative datasets can lead to unreliable predictions and simulations, potentially compromising patient safety and treatment efficacy^13,34. Additionally, the inherent complexity of AI models and DTs that leverage such algorithms as well as embedded biases in the data can limit interpretability, making it challenging to understand the rationale behind specific predictions or decisions—a crucial requirement for clinical acceptance and regulatory approval. However, this concern is less applicable to mechanistic DT models, which are based on well-understood biological and physiological processes. These models often offer greater transparency and are more interpretable, as their predictions are derived from established scientific principles rather than opaque statistical correlations. As such, mechanistic models can help bridge the gap between black-box AI systems and the need for explainable, clinically actionable insights Ensuring all models—whether AI-based or mechanistic—are trained on diverse and comprehensive datasets is therefore critical to mitigate biases and promote equitable healthcare outcomes³⁶. This includes the use of propensity score matching and inverse probability weighting when training DTs on observational datasets, ensuring balance across covariates and reducing confounding bias³⁶. Moreover, incorporating environmental and behavioral data into these models can further enhance their predictive accuracy and real-world applicability, resulting in more tailored and context-aware healthcare solutions^34,35. Lastly, regulatory frameworks must evolve alongside technological advancements, establishing rigorous standards and methodologies for validating, interpreting, and approving both AI and digital twin technologies to ensure their reliability, transparency, and clinical efficacy^9,13.

Table 1.

Advantages and disadvantages of digital twins in clinical trials: benefits, challenges, and supporting evidence

Advantages	Disadvantages
Reduced reliance on large control arms: Ethically appealing in life-threatening or rare diseases13,29,34	Potential data bias: Incomplete or non-representative datasets yield unreliable simulations13,45
Cost and time efficiency: In-silico modeling can lower expenditures and shorten trial durations34,45	Limited interpretability of complex models: Sophisticated models can limit interpretability, undermining clinician/patient trust45,46
“Fail fast” strategy: Enables earlier discontinuation of nonviable treatments, conserving resources29.	Regulatory uncertainty: No universally accepted framework for validating and approving digital twin models9,46
Personalized treatment: Integrates multi-omics and real-time data for individualized regimens29,46.	Limited integration of social and environmental factors: Excluding social determinants of health may reduce accuracy4,13
Supports adaptive trial designs: Real-time simulations guide mid-trial modifications effectively4,34	Risk of exacerbating disparities: Underrepresented groups may be overlooked if training data lack demographic diversity4,13
Enhanced Patient Safety: DTs can forecast adverse events early, allowing proactive adjustments to trial protocols and reducing patient risk21,22	Validation Framework Gaps: Lack of standardized benchmarks limits clinical validation across diseases and trial settings3,9

This table summarizes key advantages and limitations associated with the integration of digital twins (DTs) into clinical trial design. Benefits include reduced reliance on control arms, improved cost-efficiency, real-time adaptability, and enhanced patient safety. Disadvantages include data bias, limited interpretability, and regulatory and validation gaps. Citations are provided to support each claim and illustrate areas where further methodological development is needed for clinical translation.

Patient-Centered Design and Communication through LLMs

While AI-based innovations show promise in improving trial design and execution, patient engagement remains a critical factor in trial success. With policy changes such as the CURES Act in the United States, patients now have access to their medical records. However, the complexity of this information can be overwhelming. LLMs can be leveraged to turn this challenge into an opportunity. Unstructured clinical data can be parsed into more accessible, structured formats. For instance, NYUTron—designed to interpret complex EHR notes—can help match patients to the most appropriate trials³⁷. This is especially valuable given that only 2–3% of eligible candidates currently enroll in CTs³². LLMs have also demonstrated early promise in enhancing communication strategies by tailoring content to the linguistic and community preferences of potential participants, going beyond simple translation to achieve cultural adaptation. Simplifying dense consent documents can help prospective participants fully understand study procedures, risks, and benefits. Such personalization is crucial in CTs, as it helps participants fully comprehend their roles and the implications of their involvement.

These innovations also support a global trial network, boosted by decentralized methods like telehealth and remote diagnostics that reduce geographical barriers. Underpinning this shift are AI systems designed not only to personalize engagement but also to enable secure, large-scale collaboration across research institutions. As funders increasingly mandate the sharing of de-identified CT data for secondary analysis, AI-driven infrastructure plays a critical role in ensuring compliance. Federated learning models, which allow collaborative training without sharing raw data, have already been deployed in multi-institutional medical settings, enhancing diagnostic performance while preserving patient privacy³⁸. Complementary to this, synthetic data generation—particularly using generative adversarial networks (GANs)—has demonstrated high utility for clinical modeling, with privacy-preserving datasets enabling risk prediction and benchmarking³⁹. Emerging frameworks such as Federated Knowledge Recycling (FedKR) allow institutions to exchange synthetic data rather than model weights, minimizing re-identification risks⁴⁰. Moreover, in rare disease contexts, synthetic datasets have supported AI model training and virtual trials across international boundaries⁵. Together, these advances help meet funder mandates while maintaining compliance with GDPR and HIPAA.

To ensure that the benefits of LLM-driven solutions reach all patient populations safely and equitably, these technologies must be implemented alongside inclusive policies that address disparities in healthcare access. However, caution is warranted when using LLMs in patient-facing domains such as informed consent, due to ongoing concerns about the accuracy of AI-generated content, cultural sensitivity, and the risk of miscommunication. Rigorous validation, regulatory oversight, and transparency are essential for integrating these tools into clinical trial workflows responsibly.

Ethical Challenges, Trust, and Transparency

A particular concern that has been raised around AI in healthcare is the potential for AI models to perpetuate the biases in the data in regard to healthcare access and provision, while also possibly introducing new model biases and assumptions that impact clinical decisions. Addressing the “black-box” nature of AI systems is crucial, as it challenges our ability to fully understand and predict their behavior, thereby undermining trust among clinicians, researchers, and participants. To ensure quality, reliability, and ethical integrity, rigorous validation, robust regulatory oversight, and transparent algorithm development are essential. The proposed CONSORT-AI guidelines aim to enhance transparency and completeness in reporting AI-related trials by emphasizing detailed descriptions of AI systems and their integration into clinical settings. Studies highlight the need for robust methodologies to document and communicate the functioning and decision-making processes of ML models in CTs⁴¹. Transparency not only supports reproducibility but also enhances the ethical aspects of trials by ensuring all stakeholders understand underlying model assumptions and data that are driving the model outputs so that these can be considered amidst the clinical decisions.

An emerging frontier is the use of multimodal AI in CT safety monitoring. Here, heterogeneous data streams—structured EHRs, clinical notes, imaging, and wearable sensors—are fused to detect adverse events that traditional systems may miss. For example, Jimenez-Maggiora et al. describe a human-in-the-loop model in Alzheimer’s trials that combines biometric monitoring with NLP of clinical notes to flag neuropsychiatric toxicity events overlooked by conventional systems⁴². These systems offer continuous safety scoring, adaptive toxicity windows, and improved real-time reporting—critical for trials of therapies with delayed or multi-organ toxicity. To ensure clinical adoption, newer architectures like attention-based transformers and multimodal graph neural networks must be coupled with explainability overlays, to emphasize interpretable outputs for clinician trust and regulatory transparency^4,11.

Future Directions – Building a Responsible AI-Driven Trial Ecosystem

A key piece that is needed moving forward is the development of novel statistical and trial designs that enable the feasible, scalable, and rigorous evaluation of AI models and DTs. These methods must move beyond the conventional comparison of randomized treatment arms—especially in trials with limited sample sizes or ethically constrained placebo groups. Emerging approaches such as Bayesian adaptive frameworks, hybrid designs incorporating synthetic or in silico control arms, and reinforcement learning–driven protocol optimization offer promising alternatives. These designs can facilitate early detection of treatment efficacy or futility, personalize treatment paths in real time, and reduce patient burden while maintaining statistical integrity.

To realize this vision, future work must emphasize:

• Regulatory-grade uncertainty quantification.

• Robust data management that enables provenance transparency for data and models for AI and DTs

• Dynamic consent platforms using LLMs.

• Standardized pipelines for AI model reporting, such as CONSORT-AI⁴³ and DECIDE-AI⁴⁴.

• Collaborative regulatory sandboxes to trial new AI frameworks with FDA/EMA oversight.

These will collectively shift CTs from static, rigid processes to responsive, individualized systems.

Author contributions

A.B., A.M., and F.Y.M. conceptualized the Perspective. A.B. wrote the initial draft. A.M., F.M.Y., P.V., and C.C. contributed to the first draft and provided critical revisions. A.B. was responsible for data visualization, including the generation of plots and tables. All authors have read, reviewed, and approved the final manuscript.

Data availability

No datasets were generated or analysed during the current study.

Competing interests

The authors declare no competing interests.