AI-enabled clinical trials

Abstract

Randomized controlled trials remain the gold standard for evaluating the benefits and risks of new interventions, yet they face persistent challenges including high costs constraining sample sizes, complex eligibility criteria, recruitment difficulties, prolonged follow-up periods to capture hard end points, and the growing prevalence of poorly designed studies that create noise in literature. This Review proposes an artificial intelligence (AI)-enabled clinical trial engineering framework, applicable across medical specialties. This framework builds on four stages: assembling and harmonizing multimodal data to develop and validate tools for enabling trials; matching fit-for-purpose tools to the research question and target outcome; AI-supported trial conduct, including patient-to-trial matching, surrogate end points, externally matched comparator arms, digital twins, and automated data collection and curation; and faster, evidence-based go/no-go decisions that flag non-promising drugs early. Three overarching principles operate throughout: fit-for-purpose validation, continuous regulatory engagement and human oversight. This Review illustrates the framework through representative case studies, with the iBox surrogate in transplantation, the annualized relapse rate in multiple sclerosis, heart digital twins for ventricular tachycardia, AI-assisted histology (AIM-MASH) in hepatology and a large language model for patient-to-trial matching. It further maps the evolving regulatory landscape for AI-enabled clinical trials (EMA, FDA and the EU AI Act). Applied judiciously, AI-enabled clinical trials could shorten timelines, reduce costs and accelerate both the identification of effective therapies and the earlier elimination of futile ones.

Introduction

Randomized controlled trials (RCTs) are the gold standard for evaluating the benefits and risks of new interventions, offering one the highest level of evidence in medicine¹. Their primary purpose is to determine whether an intervention provides sufficient benefit relative to its risks, which serves both to advance therapeutics and to protect patients from ineffective or harmful treatments (Table 1). RCTs have changed the face of medical practice, enabling life-changing therapies across disciplines including, for example, SGLT2 inhibitors in patients with kidney disease (DAPA-CKD, NCT03036150) and heart failure (DAPA-HF, NCT03036124; EMPULSE, NCT04157751)^2,3, new targeted therapies in breast cancer (DESTINY-Breast04, NCT03734029) and lung cancer (ADAURA, NCT02511106)^4,5, or anti-amyloid monoclonal antibodies in Alzheimer’s disease (ENGAGE, NCT02477800; Clarity AD, NCT03887455)^6,7, among many.

Table 1 ∣.

Randomized controlled trials: design and definitions

Randomized controlled trials
A randomized controlled trial (RCT) is a clinical study that evaluates whether a treatment causes better outcomes than a comparator (placebo, standard of care or another treatment) Participants are assigned to groups by randomization, which helps balance known and unknown factors between groups and reduces bias Most RCTs are blinded (participants and/or investigators do not know the assigned treatment) to limit expectancy effects and biased assessment
Workflow
Drug discovery	Identify a therapeutic target and candidate compound/biologic; refine it through laboratory and preclinical studies (pharmacology, toxicology and dose exploration)
Early clinical development	First-in-human and dose-finding studies establish initial safety, tolerability and pharmacokinetics/pharmacodynamics, and help select the regimen to be tested
Design	Define the clinical question and target population (eligibility criteria), intervention and comparator, follow-up schedule, end points and data to be collected; write the protocol and statistical analysis plan
Conduct	Randomize eligible participants, initiate treatment, and collect efficacy and safety data according to the protocol, with monitoring and quality control
Analysis and reporting	Analyse results using the pre-specified statistical analysis plan and report findings transparently (including harms and limitations)
Regulatory review and approval	Submit evidence to health authorities for marketing authorization; regulators evaluate benefit–risk, end point validity and robustness of evidence
Implementation in routine care	Incorporation into clinical guidelines and real-world adoption, often complemented by post-marketing surveillance and additional studies
End points
An end point is the outcome used to evaluate the effect of the intervention. It should be clinically meaningful, measurable and defined before the trial starts
Primary end point	This is the main outcome the trial is designed to test. It drives sample size and power calculations
Secondary end point	This provides complementary evidence (such as mechanism, quality of life and intermediate outcomes) but is typically less definitive
Hard end point	This reflects direct patient benefit (such as survival, stroke, hospitalization, organ failure or functional status)
Surrogate end point	This is an earlier or easier-to-measure marker expected to predict a hard end point. It requires careful validation

Importantly, approximately 90% of drug candidates entering clinical trials fail to reach approval, most commonly due to insufficient efficacy or unacceptable toxicity⁸. This high attrition rate underscores the role of RCTs as a rigorous filter that protects patients and health-care systems from interventions whose risks outweigh their benefits. Approximately 30,000 RCTs are published each year⁹, which roughly cost US$300 billion¹⁰, reflecting their central role in biomedical progress.

Despite their relevance, RCTs face recurrent constraints. Defining optimal inclusion criteria is challenging¹¹, as overly restrictive parameters may limit generalizability whereas overly broad criteria may dilute treatment effects¹². High costs¹⁰ can constrain sample sizes, compromising studies targeting rare diseases, limiting randomization and reducing statistical power^13,14, and resulting in imbalanced groups that may undermine validity¹⁵. Years of follow-up are often required to observe hard outcomes with sufficient statistical power, such as mortality¹⁶, cardiovascular events¹⁷, cancer progression¹⁸ or organ failure¹⁹.

As a result, surrogate end points (a biomarker or intermediate outcome intended to substitute for a clinical end point, for which the change induced by an intervention is expected to correlate with, and predict, a corresponding change in the clinical end point) have been increasingly adopted to accelerate drug development, such as haemoglobin A1C in diabetes²⁰, blood pressure in cardiovascular disease²¹, proteinuria in kidney disease²², viral load in HIV²³ or progression-free survival in oncology²⁴.

However, unvalidated surrogate markers raise methodological and interpretative challenges^25,26. Their predictive performance might not capture the full complexity of disease progression and treatment response²⁷, biological specificity can be challenged by confounders acting independently of the clinical outcome²⁸, and validation is often not robust across diverse populations²⁹. Most critically, improvements in the marker frequently correlate poorly with the outcome³⁰. A marker whose treatment-induced changes fail to reliably predict treatment effects on the clinical outcome is, by definition, not a valid surrogate but a putative (candidate or unvalidated) one.

Other artificial intelligence (AI) tools and adjacent methods could address the classic shortcomings and constraints of clinical trials at other steps of their value chain. Digital twins — mechanistic, data driven or hybrid — can simulate patient trajectories and the functioning of an organ^31,32. Deeply phenotyped cohorts (in which each patient is characterized across many harmonized clinical, biological, imaging and molecular parameters, among others), by yielding mechanistic insight, can inform protocol design³³ and generate externally matched comparator arms³⁴. Agentic AI systems can assist with protocol design, patient selection, automated data collection and curation³⁵, as well as optimize and accelerate patient-to-trial matching³⁶ (Table 2).

Table 2 ∣.

AI-driven models, drugs, and trial design innovations across medical specialties

Medical Area	Role of AI	Method used	Regulatory considerations	Published or ongoing trials
Transplantation	Predicting long-term graft outcomes	iBox: prediction model integrating functional, histological and immunological parameters post-transplant to predict long-term kidney-graft survival49	EMA qualified (iBox Scoring System) (December 2022)	Validated in TRANSFORM (NCT01950819)50, BORTEJECT (NCT01873157)85, RITUX ERAH (NCT01350882)106 trials
Cardiology	Guiding ablation with MRI-based digital twin	Personalized heart digital twins	Research validation stage	Prospective study (TWIN-VT, NCT03536052)74
Pulmonology	AI-driven drug discovery	Rentosertib: AI-based target discovery and generative molecule design followed by conventional clinical development118	FDA Orphan Drug Designation (February 2023); phase IIa completed with positive results	GENESIS-IPF phase IIa (NCT05938920)118 heading to phase III
Hepatology	AI-assisted histopathological scoring for trial enrolment and end point assessment	AIM-MASH AI Assist (PathAI): ML model supporting standardized MASH CRN scoring by a single pathologist119	First AI tool qualified by EMA; First AI-powered tool qualified by FDA for MASH trials (December 2025)	Clinically validated across multiple trials, demonstrating non-inferiority or superiority to standard pathologist reads
Microbiome	AI-driven novel target discovery and molecule design	BEN-8744 (BenevolentAI): Benevolent Platform knowledge graph with ML-based gene prioritization identified PDE10 as a novel therapeutic target for ulcerative colitis, leading to the development of a first-in-class oral PDE10 inhibitor120	Early clinical development stage	Phase I RCT completed with favourable safety, tolerability, and PK profile120; phase Ib/IIa in patients with moderate-to-severe UC in preparation
Infectious disease	AI-driven drug repurposing/indication discovery	Baricitinib (BenevolentAI): Benevolent Platform knowledge graph used to rapidly identify baricitinib, an approved JAK inhibitor, as a candidate therapy for COVID-19121,122	FDA approval (May 2022)123	Validated in phase III RCT and COV-BARRIER trial (NCT04421027); associated with reduced mortality and endorsed by WHO124
Autoimmune disease	AI-driven prediction of disease activity and treatment response	Rheuma Care Manager (RCM): AI-powered CDSS with machine learning ensemble flare prediction model to predict individual flare risk125,126	No FDA/EMA approval so far	Flare prediction model developed on RETRO phase III RCT (NCT02779114)127; CDSS pilot evaluation125; PATIO multicentre RC (Dutch Trial Register NL9798, results pending)128
Oncology	AI-designed immuno-oncology molecule	EXS-21546: A2A receptor antagonist designed using Exscientia Centaur Chemist generative AI platform129	Clinical development ongoing	Phase I/II trial in renal cell carcinoma and non-small cell lung cancer (NCT04727138 phase I; NCT05920408 phase Ib/IIa)
Neurology	AI-generated digital twins for trial efficiency	Unlearn.AI digital twin generators: ML models trained on historical data to predict individual placebo trajectories72,130	FDA and EMA aligned with methodology72	AWARE trial (tilavonemab, early Alzheimer’s disease, phase II; NCT02880956)72
Ophthalmology	AI-based automated diabetic retinopathy screening	IDx-DR/LumineticsCore: autonomous AI diagnostic system for more-than-mild diabetic retinopathy131	FDA clearance: first autonomous AI diagnostic system in any field of medicine	Validated in a pivotal multicentre prospective trial (NCT02963441) and real-world implementation studies showing high diagnostic performance and expanded screening access132,133
Cross-specialty	AI-derived prognostic covariate for trial design	PROCOVA (Unlearn.AI): digital twin-derived prognostic score used as covariate adjustment; reduces sample size 20–40% or increases power134	EMA qualified (September 2022); FDA confirmed alignment with guidance	Applied to AWARE trial (tilavonemab, early Alzheimer’s disease, phase II; NCT02880956)72; prospective applications in other therapeutic areas under evaluation

Examples were selected to illustrate different applications of AI in clinical trial design across key medical specialties. Inclusion aimed to represent distinct roles of AI (surrogate end points, mechanistic digital twins, AI-driven drug discovery and drug repurposing) and varying stages of regulatory maturity (from research validation to FDA/EMA qualification). Examples were identified through targeted searches of PubMed, ClinicalTrials.gov, regulatory databases (FDA and EMA) and company press releases (Supplementary information). AI, artificial intelligence; CDSS, clinical decision support system; ML, machine learning; PK, pharmacokinetics; RCT, randomized controlled trial; UC, ulcerative colitis.

Together, these tools could make trials faster, more efficient and more informative across recruitment, conduct and analysis. However, like surrogate end points, none of these tools can be adopted uncritically: each must be demonstrated to be fit for its intended purpose before it can be accepted by regulator and reliably integrated into trial design.

Therefore, in this Review, we propose an AI-enabled clinical trial engineering framework. In this framework, AI components are inserted into the conventional trial workflow. It spans (1) assembling and harmonizing multimodal data to develop and validate tools for enabling trials: validated surrogate end points, digital twins and agentic systems; (2) matching fit-for-purpose tools to the research question and target outcome; (3) AI-supported trial conduct, including patient-to-trial matching, externally matched comparator arms, prognostic prediction and data monitoring, and curation; and (4) faster, evidence-based go/no-go decisions that flag non-promising drugs early. Three overarching principles operate throughout: fit-for-purpose validation, continuous regulatory engagement and human oversight (Fig. 1).

Fig. 1 ∣. The AI-enabled clinical trial engineering framework.

Artificial intelligence (AI) and adjacent tools are inserted into the conventional trial workflow to improve each step. The left part shows tool development, spanning validated surrogate end points, digital twins and agentic AI. The right part shows the conventional trial workflow, from the research question through protocol and trial design, site selection and patient recruitment, and the on-treatment phase, leading to faster go/no-go decisions. LLM, large language model.

This Review focuses primarily on drug and biologic development. However, the principles discussed are not specific to drugs and could generalize to other interventions or regulated domains where trials are conducted. Furthermore, the principal context of this Review is pre-market development; however, consequential applications of AI also lie in the post-market setting, where large-scale data curation and standardized outcome measurement could improve the assessment of comparative effectiveness, value, long-term toxicity and real-world outcomes. Importantly, the same approach developed for pre-market trials can be repurposed for these questions.

Challenges in the design and conduct of clinical trials

Operational and methodological barriers

Clinical trials face persistent operational barriers. Recruitment is a major challenge: less than half of trials meet their pre-specified enrolment target upon completion³⁷, with complex eligibility criteria, geographic constraints and participation burden contributing to high screen-failure rates¹². Patient dropout further decreases statistical power, particularly in trials requiring prolonged observation³⁸. These inefficiencies — compounded by labour-intensive manual data collection, suboptimal site selection and costly monitoring processes — translate directly into longer timelines and delayed identification of both beneficial and harmful interventions.

These difficulties are intrinsic to rigorous evaluation and should not be taken as a justification for adopting less valid methods of assessing risk and benefit. By contrast, the resolve to conduct adequately designed and sized trials should be strengthened wherever such trials are feasible. The purpose of the AI-based approaches is to help conduct rigorous trials more efficiently.

Moreover, the increasing number of poorly designed studies with unreliable data, known as zombie trials³⁹, creates noise in scientific literature. The scale of this problem is considerable: analyses have identified patterns of fabricated individual patient data in a substantial proportion of trials submitted to major journals⁴⁰. The emergence of AI-generated content has amplified this threat, as fabricated or low-quality manuscripts can now be produced at industrial scale by paper mills leveraging generative AI, further straining the capacity of peer review and meta-analytic synthesis to distinguish signal from noise⁴¹. However, the data-curation and anomaly-detection components of an AI-enabled trial pipeline can conversely be used to screen individual-patient data for implausibility and strengthen evidence integrity.

The surrogacy challenge of correlation versus substitution

A critical distinction in clinical trial design separates three concepts: a prognostic biomarker (one that predicts outcome irrespective of treatment and is used for risk stratification and enrichment of high-risk populations), a predictive biomarker (one that identifies patients more likely to benefit from a specific treatment and is used as companion diagnostics or for subgroup identification) and a true surrogate end point (one for which the change induced by treatment reliably predicts the change in the clinical outcome). History offers cautionary examples: anti-arrhythmic drugs that suppressed premature ventricular contractions yet increased mortality in the CAST trial (NCT00000526)⁴² or the CETP inhibitor torcetrapib that raised HDL cholesterol yet increased cardiovascular events and death in the ILLUMINATE trial (NCT00134264)⁴³.

Importantly, surrogate end points must be distinguished from direct clinical outcomes. The basis for regulatory approval remains whether an intervention helps patients feel better, function better or live longer. Symptomatic and functional outcomes are themselves valid, patient-centred clinical end points that can support approval in their own right. Biomarkers, by contrast, fall outside this ‘feel–function–survive’ paradigm: they acquire regulatory relevance only when they are formally linked to such outcomes or qualified as substitutes for them through the surrogacy criteria described above.

In addition, biomarker-based surrogates should also be distinguished from intermediate clinical outcomes, which are clinical events that patients directly experience and that lie on the causal pathway to the long-term end point. For instance, the annualized relapse rate (ARR) — the number of confirmed clinical relapses per patient-year — is the principal efficacy measure in relapsing–remitting multiple sclerosis; it is both a clinically meaningful intermediate outcome and a metric validated at the trial level as a surrogate for longer-term disability progression, which is why it anchors regulatory approvals in the field.

Safety, time-dependent benefit–risk dynamics and the post-market evidence gap

Clinical trials serve also to characterize safety. Off-target effects may emerge on timescales different from those captured by surrogate end points⁴³. Moreover, benefit–risk curves can cross over time: an early hazard may convert to a late benefit, or vice versa¹⁶. A telling example is provided by renin–angiotensin system inhibitors in heart failure: blood pressure and renal-function measures may transiently move in an apparently unfavourable direction shortly after initiation, whereas the clinical benefit emerges only later (SOLVD, NCT00000516)⁴⁴. Abandoning these therapies based on short-term biomarker movement would have been an error. This example shows that the temporal relationship between a biomarker and a clinical outcome is itself part of what must be validated. Shortening follow-up through surrogate-based projections therefore risks capturing only one phase of this dynamic³⁰. These considerations do not invalidate AI-based acceleration but they impose clear safeguards: surrogate-anchored trials must be complemented by structured post-trial surveillance and systematic pharmacovigilance⁴⁵. Importantly, the same AI infrastructure — cohorts, digital twins and real-world databases — can be repurposed for post-marketing studies addressing comparative effectiveness, treatment duration and long-term safety⁴⁶.

Moreover, regulatory approval does not resolve all clinically relevant questions: comparative effectiveness, cost–effectiveness, optimal treatment duration, and real-world value across diverse populations require continued investigation after market entry. AI-enabled trials are not intended to replace comprehensive post-approval evaluation, but to accelerate the initial efficacy and safety signal. The ambition is thus twofold: faster access to effective therapies and earlier identification of futile or dangerous ones.

In addition, wherever an adequately powered trial with relevant follow-up is feasible, it remains the preferred option. When this is not feasible and no validated surrogate is available, decisions rest on a field-level consensus about what reduces uncertainty most feasibly and what level of residual uncertainty is acceptable given the potential for benefit. This is the logic underlying accelerated-approval pathways, in which an unvalidated surrogate must be ‘reasonably likely’ to predict clinical benefit.

Qualifying surrogate end points

Why robust surrogates are both necessary and difficult

Developing, validating and ultimately qualifying robust surrogate end points for long-term outcomes is complex. It requires building large, deeply phenotyped cohorts with plausible links to disease mechanisms, strong individual-level and trial-level associations with hard outcomes, and high prediction performances across diverse clinical scenarios and patient subpopulations.

The clinical pressures that motivate this work differ by field but share a common need for regulator-grade evidence. For instance, oncology must anticipate survival and durable benefit within highly heterogeneous tumours and combination regimens. Neurology needs sensitive and validated proxies for very long-term functional decline in conditions with slow trajectories. Cardiology must predict composite events in the context of rapidly evolving standards of care and background therapies. Transplantation faces the dual pressure of a growing global organ shortage⁴⁷ and increasing prevalence of patients suffering from end-stage organ disease⁴⁸, and therefore maximizing allograft longevity has emerged as a key objective. Across all areas, transparent methodology, external validation and demonstration of clinical utility are prerequisites for regulatory acceptance.

Development of integrative digital biomarkers

A central bottleneck in kidney transplantation is that the key end point of death-censored allograft failure accrues over 10–20 years, making trials of new immunosuppressants impractically long and underpowered. In an effort to develop predictive models for kidney allograft failure, the iBox marked a step forward in the field of transplantation⁴⁹. After building deeply phenotyped cohorts of kidney recipients and considering a large series of candidate parameters associated with allograft failure, eight determinants were identified, including two functional biomarkers, namely estimated glomerular filtration rate and proteinuria; one immunological biomarker, namely the mean fluorescence intensity of anti-HLA donor-specific antibodies; and four histological parameters capturing active and chronic allograft lesions. This integrative digital biomarker model projected long-term outcomes in trials⁵⁰, demonstrated strong discrimination capability in paediatric settings (c statistic of 0.81)⁵¹, surpassed clinicians in forecasting allograft failure⁵², and demonstrated comparable or superior performance to other methods, including machine learning models and competing risks model^53,54 (Supplementary Fig. 1). Its design and reliability have gained regulatory recognition, including endorsement by the Transplant Therapeutics Consortium and qualification as a secondary efficacy end point by the EMA. The generalizable lesson is that a multivariable prognostic model built on standardized inputs can achieve regulatory qualification, offering a template for turning a prediction model into a usable trial end point.

In oncology, the FDA has approved progression-free survival and objective response rate as surrogates for overall survival in specific cancer types, such as advanced ovarian cancer⁵⁵, HER2-positive and HER2-low metastatic breast cancer⁵⁶ and metastatic colorectal cancer⁵⁷. Scientific professional societies also endorse the use of surrogate to assess value of drugs. For instance, the European Society of Medical Oncology has included progression-free survival in its ‘magnitude of clinical benefit scale’⁵⁸. Although the use of surrogates is controversial in diseases with very poor outcome and for drugs with modest efficacy, they become more relevant when drugs are highly effective in a group of patients with a life expectancy of several years.

In neurology, although various end points have been explored for multiple sclerosis treatments since the 1990s, the reduction in ARR is the primary surrogate for relapsing-remitting multiple sclerosis therapies, underpinning FDA approvals for drugs such as interferon-β1a and fingolimod^59–61. More recent examples include ocrelizumab⁶², with meta-analyses confirming the robustness of ARR in predicting disability progression⁶³, although correlations with long-term patient-centred outcomes remain moderate⁶⁴. In Alzheimer’s disease trials, amyloid PET imaging changes — measuring beta-amyloid plaques via tracers such as florbetapir — gained acceptance as pharmacodynamic biomarkers for target engagement since the early 2010s through initiatives such as the Alzheimer’s Disease Neuroimaging Initiative⁶⁵. These efforts supported the FDA accelerated approvals for aducanumab in 2021 (EMERGE, NCT02484547; ENGAGE, NCT02477800) and lecanemab in 2023 (Clarity AD, NCT03887455) with plaque reductions on Centiloid scales linked to slowed decline⁷.

However, their correlation with clinical benefits such as the Clinical Dementia Rating–Sum of Boxe scores remain under evaluation due to incomplete capture of factors such as tau pathology⁶⁶.

Digital twins and tools for trials

Not all tools discussed here are AI in the strict sense — although the boundary between genuine AI and classical statistical methods is itself subject to discussion^67,68 — but are adjacent methods, such as classical surrogate end points (ARR), statistical prognostic modelling, mechanistic simulation (heart digital twins), digital monitoring or target-trial emulation.

From surrogate end point to surrogate patient

When a validated model predicts a patient’s long-term outcome shortly after treatment initiation, it effectively creates a digital twin of that patient within the trial context. In this sense, a robust surrogate end point becomes a surrogate patient trajectory that complements other measures and enables earlier and more reliable projections at both the patient and the trial level (Fig. 1).

Distinct definitions of digital twin.

The concept of digital twins has gained substantial traction across medical specialties, although the term encompasses methodologically distinct approaches that should not be used interchangeably. Mechanistic digital twins are physics-based — or physiology-based — patient-specific simulations grounded in biophysical principles and are primarily suited to intervention planning, in silico experimentation and trial enrichment⁶⁹. Data-driven digital twins are learned from cohort-level data to estimate individualized trajectories or risks and are therefore better suited to end point projection, prognosis and covariate adjustment^49,70. Hybrid digital twins combine mechanistic structure with data-driven learning and can potentially support both functions⁷¹. As these categories differ in their assumptions, sources of uncertainty, and intended use, they require distinct validation pathways, as discussed in the validation framework section below.

In neurology, digital cognitive twins represent real-time, personalized AI models of cognitive states, continuously updated through wearables and biometric data to simulate therapeutic responses and detect early decline. For example, in the AWARE trial (NCT02880956), tilavonemab did not demonstrate clinical efficacy in progressive supranuclear palsy⁷²; however, the digital twin methodology itself performed as intended, generating individual-level placebo predictions that closely matched observed trajectories.

Cardiology in particular has advanced this technology with AI-accelerated digital twin modelling⁷³ (that reduces computation time from hours to seconds), and clinical measurements-validated digital twin-guided ablation therapy that enables precise catheter targeting with improved efficacy compared with traditional approaches^74–77. This approach is now being evaluated prospectively in the TWIN-VT study (NCT03536052), the first clinical trial of digital twin-guided ablation conducted under an FDA Investigational Device Exemption. In addition, heart digital twins create patient-specific virtual replicas based on clinical imaging and other data, simulating the diseased organ’s functioning while integrating physics-based mechanistic insights across multiple spatial scales^78,79. Genotype-specific digital twins further enhance precision by incorporating genetic variant information with structural remodelling patterns to predict ventricular tachycardia circuits⁸⁰.

Mechanistic insight and design optimization

The same deeply phenotyped cohorts and harmonized databases that underpin surrogate development can be mobilized pre-launch to map biomarker–outcome associations; identify treatment-responsive subgroups; refine inclusion/exclusion, prototype stratification and interim rules; and stress-test end point behaviour for sensitivity and generalizability⁸¹. In this view, AI functions as a design assistant that converts mechanistic and multimodal evidence into concrete eligibility and stratification rules, enabling both prognostic enrichment (prioritizing higher baseline risk for event accrual and efficiency) and predictive enrichment (selecting biology-defined responders). Alternative schemas can be simulated to quantify effects on sample size, power and treatment heterogeneity, and assess fairness and transportability across sites and demographics.

In transplantation, unbiased biopsy cohorts uncovered complement-driven vascular injury⁸², natural killer cell cytotoxicity (missing-self)⁸³, microvascular inflammation⁸⁴ and plasma-cell-mediated antibody activity⁸⁵ – mechanisms that refined classifications and explained divergent outcomes. This paradigm now guides trials beyond transplantation, testing complement inhibitors⁸⁶, anti-CD38 agents⁸⁷, Fc-pathway modulators⁸⁸ and repurposed drugs within mechanistically defined phenotypes⁸⁹. Such frameworks enable enriched selection, mechanistic end points and earlier detection of signals that clinical features miss.

In oncology, several programmes aim at building a biological twin of the patient that recapitulates the biology of the disease and the host. The first generation of this programme is illustrated by the SAFIR programme^90,91. In these trials (SAFIR01/UNICANCER, NCT01414933; SAFIR02-Breast, NCT02299999; SAFIR02-Lung, NCT02117167), genomic profiling of the tumour was conducted to identify mechanisms of cancer progression in each patient and match the treatment accordingly. Genomic profiling was interpreted by a human based on available knowledge and frameworks. In the new versions of the precision medicine trials in oncology, the molecular profile of the digital twin is multimodal and is interpreted by AI agents (for example, the PRISM/MEDITWIN initiative and the Gustave Roussy-Aitia). Another example is the CAPRI trial (NCT07449923), a phase III RCT in patients receiving oral anticancer therapy, in which a digital remote-monitoring tool combined with nurse-navigator intervention reduced hospitalizations⁹².

In addition, modern biomedical research is shifting from clinical or histologic labels to biological programmes revealed by multimodal, multiomics technologies, reshaping trial design across oncology, cardiology, neuroscience and immunology. Integrating histology, multiplex immunostaining, gene-expression signatures and single-cell/spatial transcriptomics yields mechanistic tissue maps that expose disease archetypes invisible to conventional end points.

LLMs for patient-to-trial matching

Patient recruitment is a leading cause of trial delay and failure, and manual screening of records against complex eligibility criteria is slow and labour intensive. Large language models (LLMs) now enable end-to-end, zero-shot patient-to-trial matching that materially accelerates recruitment³⁶. In recent evaluations, LLM pipelines combining retrieval, criterion-level eligibility classification, and trial-level ranking recalled most relevant trials from large corpora, achieved expert-level accuracy on patient–criterion decisions with faithful rationales, and produced trial rankings that correlated strongly with human judgements⁹³. These systems also reduced screening time, suggesting immediate operational impact. Embedding such LLM matchers into pre-screening and site-feasibility workflows can shrink screen-failure rates and focus coordinator effort on the most promising candidates⁹⁴. However, practical failure modes must be acknowledged, including hallucinated eligibility criteria, sensitivity to clinical note quality and performance degradation with non-English records or non-standardized documentation^95,96.

Generation of externally matched comparator arms

For new clinical trials, deeply phenotyped cohorts can also support externally matched comparator arms, which are control groups assembled by matching each trial participant to closely similar patients drawn from a curated external cohort with known outcomes. For example, in a two-arm study, trial participants can be matched to comparable patients drawn from a curated cohort using clinical, biological, immunological, histological or imaging criteria, thereby constructing an externally matched comparator arm with known longitudinal outcomes⁹⁷. This strategy becomes particularly valuable when standard-of-care arms are impractical, underpowered or ethically constrained (Fig. 1).

However, matching on observed covariates does not eliminate all residual confounding. Exchangeability assumptions must be explicitly stated and tested, temporal alignment is essential to avoid immortal time bias and sensitivity analyses for unmeasured confounding should be standard practice. The target trial emulation framework (which specifies the protocol of the ideal randomized trial and then reproduces each of its elements in observational data⁹⁸) provides a rigorous methodological structure for these designs, and the FDA has issued specific guidance on the design and conduct of externally controlled trials. When these standards are met, externally matched comparators offer a valuable complement to randomized designs.

Towards an AI-enabled clinical trial engineering framework

The tools discussed in the preceding sections are components of an overarching framework whose ambition is to improve the design and conduct of clinical trials. We refer to this as the AI-enabled clinical trial engineering framework: rather than replacing the conventional trial, AI components are inserted into its existing workflow to make each step faster, more informative and more efficient (Fig. 1 and Table 3).

Table 3 ∣.

Advantages and challenges of using AI in clinical trials

Topics	Advantages	Challenges	Validation / mitigation
Protocol design	More efficient protocol design: optimized eligibility, stratification and end point selection from deeply phenotyped cohorts; scenario simulations for accrual/power	Data quality and governance: harmonization, missingness, shift/bias control, data lineage, access permissions and reproducibility across sites and time	Pre-registration of design choices; documented data lineage; bias and missingness audits; reproducibility checks across sites and time
Recruitment and site selection	Improved recruitment and site selection through predictive feasibility and diversity targeting; reduced screen-failures and better representativeness	Bias, fairness and generalizability risks if training data are unbalanced or context limited; risk of spurious correlations and performance drift in new settings	Fairness and subgroup-performance audits; external and temporal validation; monitoring for performance drift in new settings
External comparators	External/matched real-world comparators that reduce control-arm burden and increase efficiency when standard arms are impractical or unethical	Regulatory and ethical constraints for real-world data use (consent, privacy and data provenance), plus clear justification that the matching of external controls is of quality	Target-trial emulation with explicit exchangeability assumptions; sensitivity analyses for unmeasured confounding; quality metrics for external matching
Digital twin and early prediction of findings	Earlier, patient-level risk projections (digital-twin trajectories) to inform interim decisions and adaptive designs, potentially shortening follow-up	Operational change management: integrating AI into SOPs, investigator training, and stakeholder acceptance; avoiding over-reliance on algorithmic outputs	Prospective validation against observed outcomes; calibration assessment; human-in-the-loop review to avoid over-reliance on algorithmic outputs
Multimodal integration	Higher signal-to-noise via multimodal integration (clinical, biological, immunological, histological, imaging and omics) and continuous digital measures; improved end point precision	Technical debt and infrastructure costs: secure computing, interoperability, MLOps for versioning and audit trails; cybersecurity for sensitive health data	Analytical and clinical validation of each modality; MLOps with versioning and audit trails; cybersecurity and data-governance controls
Standardized reporting	Standardized, auditable reporting — auto-generated protocol/SAP text, model cards, CONSORT-AI appendices — that improves transparency and reproducibility	Legal and compliance complexity across regions (FDA/EMA and local privacy rules); aligning documentation with evolving guidance and audits	Adherence to CONSORT-AI/SPIRIT-AI; model cards and version control; independent audit of auto-generated protocol/SAP text
Resource optimization	Resource optimization and cost containment: shorter timelines, fewer futile arms, faster ‘no-go’ decisions and earlier identification of responders	Multidisciplinary expertise (clinicians, statisticians, ML engineers and data stewards) and sustained funding to maintain quality over time	Pre-defined go/no-go criteria; independent statistical oversight; sustained multidisciplinary governance to maintain quality over time

AI, artificial intelligence; ML, machine learning; MLOps, machine learning operations; SAP, statistical analysis plan; SOP, standard operating procedure.

Concretely, this framework unfolds across four stages. It begins by assembling and harmonizing multimodal data to develop and validate fit-for-purpose tools (validated surrogate end points, digital twins and agentic systems). Each tool is then matched to a specific research question and target outcome, before being deployed to support trial conduct (from to patient-to-trial matching to externally matched comparator arms, to prognostic prediction and data collection and curation). Ultimately, these components converge on faster, evidence-based go/no-go decisions. Underpinning all four stages are three overarching principles: fit-for-purpose validation, continuous regulatory engagement and human oversight.

Agentic AI as the operating component of clinical trials

LLMs and tool-using agents act as the orchestration layer of the AI-enabled clinical trial engineering framework. In this context, ‘agentic’ refers to AI assistants that can use tools, access databases and chain multistep tasks under investigator supervision.

A scoping review of agentic AI in health care found that the majority of studies were exploratory or technical, with limited real-world validation and no demonstrated clinical benefit in RCTs so far⁹⁹. This early-stage evidence underscores that agentic AI for clinical trial remains an emerging application that must be developed with appropriate caution.

We thus propose to distinguish three levels of maturity: (1) currently implemented tools include LLM-based patient-to-trial matching, automated data extraction and structuring from clinical notes, and protocol drafting assistance with regulatory checklists; (2) emerging applications include real-time data quality monitoring, adaptive trial monitoring with drift detection and mechanistic analysis for end point refinement — applications for which proof-of-concept evidence exists but prospective validation is ongoing; (3) future applications include fully orchestrated trial design agents that simulate accrual, power and end point behaviour under multiple scenarios, assemble matched real-world comparators, and draft complete protocol and statistical analysis documents. These remain forward-looking concepts.

At every level, the following practical limitations must be acknowledged: hallucination risk (generation of plausible but incorrect outputs), reproducibility challenges across model updates, auditability requirements for regulatory compliance and the absolute necessity of human-in-the-loop oversight for all high-impact decisions. The orchestration layer is essential to make the system usable, transparent and compliant.

Fit-for-purpose validation across AI applications

Before a tool can be used in a clinical trial, its fitness to a specific purpose must first be demonstrated. Different AI applications require distinct validation standards that align with the established ACCE framework (analytic validity, clinical validity, clinical utility and ethical, legal and social implications)¹⁰⁰ and guidance on performance evaluation of predictive AI models^101,102.

Analytical validity asks whether the model accurately measures what it claims to measure, encompassing discrimination, calibration and overall performance. Clinical validity asks whether outputs correlate with clinically meaningful outcomes in external, diverse populations, with attention to transportability and fairness across subgroups. Clinical utility asks whether the model improves decisions compared with standard practice. For surrogate end points, this demands trial-level evidence that treatment effects on the surrogate predict treatment effects on the hard outcome^27,29; for digital twins, prospective comparison with standard care; for LLM-based tools, demonstration of operational gains. Regulatory fitness-for-purpose asks whether the model meets the evidentiary bar for its intended context of use, which varies across EMA qualification pathways, the FDA credibility framework and EU AI Act requirements for high-risk systems.

The stringency of these requirements differs by application. Surrogate end points require the full path from accurate calibration through external validation to trial-level treatment-effect capture and formal regulatory qualification. Mechanistic digital twins demand physics-based verification and prospective clinical comparison. Data-driven digital twins require cross-cohort generalization and demonstrated efficiency gains. Externally matched comparators call for rigorous matching metrics, sensitivity analyses for unmeasured confounding and target trial emulation. LLM-based tools require clinically grounded evaluation using expert benchmarks and evidence of real-world operational impact before deployment¹⁰³. Although the FDA has issued draft guidance for AI-enabled devices, no LLM is currently authorized by the FDA as a clinical decision support device in this form¹⁰⁴. Agentic AI in health care remains at a very early stage, with the current literature dominated by exploratory or technical studies, limited real-world validation and no demonstrated randomized-trial benefit to date⁹⁹.

Beyond initial validation, deployed AI systems require ongoing governance throughout their life cycle. This includes dataset lineage (full traceability of data sources and preprocessing steps), version control (documented tracking of every model iteration with corresponding performance benchmarks), audit trails (immutable records of all AI-assisted decisions and outputs) and continuous performance drift monitoring. These are prerequisites for regulatory compliance and stakeholder trust.

Surrogate end points across therapeutic areas

AI-enabled clinical trials are moving from concept to practice. The predictive accuracy of a qualified allograft failure prediction system has been validated through several landmark trials, including TRANSFORM (NCT01950819)¹⁰⁵, which investigated the clinical relevance of everolimus (an mTOR-inhibitor immunosuppressant), BORTEJECT trial (NCT01873157)⁸⁵ and RITUX ERAH (NCT01350882)¹⁰⁶, which evaluated new drugs for late and acute antibody-mediated rejection, respectively. In addition, in a cohort study examining post-transplant belatacept conversion¹⁰⁷ — a CTLA-4-Ig costimulation blocker used as an alternative to calcineurin inhibitors to reduce their nephrotoxicity — the same prediction system confirmed equivalent allograft failure risk between matched treatment groups at conversion time, effectively simulating RCT conditions. This model is now actively shaping new clinical trials across multiple therapeutic approaches.

In neurology, disability accrual (measured by the Expanded Disability Status Scale) unfolds over years, so trials powered directly on disability are long and large. As such, the ARR provides an earlier readout. Its predictive accuracy, when used as a surrogate end point for multiple sclerosis treatments, has been validated through landmark trials, including PRISMS for interferon-β1a⁵⁹; FREEDOMS¹⁰⁸ (NCT00289978) and TRANSFORMS⁶⁰ (NCT00340834) for fingolimod; and OPERA I (NCT01247324) and II (NCT01412333) for ocrelizumab⁶², demonstrating ARR reductions correlating with reduced MRI lesion activity and disability progression. In meta-analyses, ARR outperformed other metrics (such as MRI lesion counts and short-term expanded disability status scale change) in predicting the worsening of sustained expanded disability status scale¹⁰⁹. ARR is now shaping new clinical trials, with companies integrating it as a primary or secondary end point in ongoing programmes, including of atumumab¹¹⁰, ocrelizumab⁶² and fenebrutinib. Importantly, the limitations of ARR are recognized: it does not fully capture disability progression independent of relapses, and its surrogacy is weaker in progressive phenotypes. The generalizable lesson is that the acceptance bar for any surrogate is trial-level evidence that the treatment effect on the surrogate predicts the effect on the hard outcome.

In cardiology, identifying ablation targets in ventricular tachycardia is empirical and invasive, with variable outcomes. The TWIN-VT study (NCT03536052) prospectively tested personalized digital twin-guided ablation for ischaemic ventricular tachycardia in ten participants. Ventricular tachycardia was rendered non-inducible in all patients after ablation of digital twin targets, and eight of ten participants remained disease free without anti-arrhythmic drug therapy, with no periprocedural complications¹¹¹. These results demonstrate the feasibility and clinical promise of mechanistic digital twin technology for guiding interventional procedures. The digital twin is reconstructed from late-gadolinium-enhancement cardiac MRI integrated into biophysical electrophysiology simulations. This approach creates a mechanistic, physics-based twin distinct from purely data-driven prediction model. Its limitations include computational demands, dependence on imaging quality and modelling assumptions. The generalizable lesson is that mechanistic digital twins can convert imaging into actionable, patient-specific predictions, but must meet prospective clinical validation before broad adoption.

In hepatology, histological scoring of metabolic dysfunction-associated steatohepatitis — steatosis, inflammation, hepatocyte ballooning and fibrosis — is subjective, and high inter-reader variability inflates sample size and can obscure treatment effects. The AIM-MASH AI Assist aims to address this bottleneck. It is a machine learning, computer-vision model that scores digitized liver biopsies to standardize MASH CRN grading and enable reproducible single-reader assessment. It as clinically validated across trial datasets, showing agreement with expert consensus and improved reproducibility relative to standard pathologist reads. The AIM-MASH AI assist became the first AI tool qualified by both the EMA and FDA for clinical trials. Its limitations include dependence on scan and staining quality and the need to monitor cross-laboratory generalization. The generalizable lesson is that where an end point is reader dependent, AI standardization can directly improve trial efficiency and follows a clear, context-of-use regulatory pathway.

Regulatory landscape and ethics

The regulatory framework governing traditional clinical trials is well established, built upon harmonized international standards including ICH Good Clinical Practice, the EU Clinical Trials Regulation and FDA guidance documents (Table 4). In parallel, a rapidly evolving regulatory landscape is emerging specifically for the use of AI in clinical trial design and drug development (Table 5). Key milestones include the FDA’s risk-based credibility assessment framework for AI models supporting regulatory decisions, the EMA’s reflection paper on AI in the medicinal product life cycle, and the joint FDA–EMA guiding principles for good AI practice in drug development, which establish a shared transatlantic foundation for AI-enabled trials. The EU AI Act, the world’s first comprehensive AI legislation, further classifies AI systems used in medical devices as high risk, imposing requirements for transparency, human oversight and conformity assessment. Notably, the USA has not enacted equivalent comprehensive federal AI legislation, relying instead on sector-specific FDA guidance.

Table 4 ∣.

Regulatory landscape of traditional clinical trials

Regulatory body	Document	Scope and key provisions
European Medicines Agency	Clinical Trials Regulation (2022, final)	Current EU regulatory basis for all clinical trials. Introduction of a single EU portal for multinational trial submission, harmonized assessment, streamlined safety reporting, enhanced transparency and risk-proportionate approach for low-intervention trials
Food and Drug Administration	Conducting Clinical Trials with Decentralized Elements (2024, final)	Recommendations for implementing decentralized trial elements (telehealth and home-based activities)
	Enhancing Participation in Clinical Trials — Eligibility Criteria, Enrollment Practices, and Trial Designs (2025, final)	Recommendations to broaden eligibility criteria, avoid automatic exclusions of higher-risk populations (older and comorbid patients) and design trials reflecting real-world patient populations
	Clinical Trials Guidance Documents (2025, final)	Current US regulatory basis for all clinical trials. Recommendations on the conduct of clinical trials, good clinical practice and human subject protection
International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use	Statistical Principles and Estimands and Sensitivity Analysis (2020, final)	Framework for defining treatment effects through estimands (population, variable, intercurrent events and summary measure), aligning trial objectives with statistical analysis
	General Considerations for Clinical Studies (2022, final)	Promotes stakeholder engagement throughout the trial life cycle and provides a structured approach for efficient trial design focused on elements most critical to achieving reliable results
	Good Clinical Practice (2025, final)	Introduces risk-based quality management, explicit data governance requirements, support for decentralized trial elements, e-consent and digital health technologies

This table summarizes the current regulatory frameworks governing the design, conduct and reporting of randomized controlled trials, including international harmonized standards (ICH), European Union regulations and Food and Drug Administration guidance documents.

Table 5 ∣.

Regulatory landscape of AI-enabled trials

Regulatory body	Document	Scope and key provisions
Food and Drug Administration and European Medicines Agency	Guiding Principles of Good AI Practice in Drug Development (2026, final)	Joint transatlantic framework covering AI use across the medicines life cycle; emphasizes human-centric approach, risk proportionality, data quality, multidisciplinary expertise and transparent model development. Establishes a shared regulatory foundation for sponsors conducting multinational AI-enabled trials
European Medicines Agency	Reflection paper on the use of AI in the medicinal product life cycle (2024, final)	Provides considerations for AI use in clinical trial design, patient selection, end point assessment and manufacturing
	Qualification of Novel Methodologies programme for Medicine Development (ongoing)	Formal regulatory qualification opinions for innovative AI-based tools and AI-derived end points
Food and Drug Administration	Considerations for the Use of AI to Support Regulatory Decision-Making for Drug and Biological Products (2025, draft)	Risk-based credibility assessment framework for AI models used to generate data supporting drug safety, effectiveness or quality decisions
	Center for Clinical Trial Innovation program (ongoing)	Dedicated pathway for sponsors to engage FDA early on the use of AI in clinical trial design, including novel digital end points and AI-augmented monitoring
	Drug Development Tools Qualification Program (ongoing)	Formal pathway to qualify novel methodologies as drug development tools. Enables cross-trial qualification of AI-based surrogate end points and prognostic scores
European Union	Artificial Intelligence Act (2024, final)	World’s first comprehensive AI legislation. Risk-based classification (unacceptable, high, limited or minimal). AI systems in medical devices classified as high-risk requiring transparency, human oversight, data governance and conformity assessment
Medicines and Healthcare products Regulatory Agency (UK)	AI regulatory strategy and AI Airlock sandbox (2026, final)	Principles-based regulation aligned with five pillars (safety, transparency, fairness, accountability and contestability)

This table summarizes the emerging regulatory frameworks governing the use of AI in clinical trial design and drug development, including AI-specific guidance from the FDA and EMA, the EU Artificial Intelligence Act, joint transatlantic initiatives and the UK Medicines and Healthcare products Regulatory Agency AI regulatory strategy. AI, artificial intelligence.

In addition, the deployment of AI in clinical trial design raises specific ethical, safety and data governance challenges. When AI systems influence trial’s critical decisions — such as patient selection, end point assessment or treatment allocation through digital twins — participants must be adequately informed, as recommended by the SPIRIT-AI reporting extensions⁴⁵. AI models trained on non-representative datasets risk perpetuating health disparities: both the FDA and EMA emphasize the need for proactive bias audits and demographic performance monitoring, whereas the EU AI Act mandates bias mitigation for high-risk systems. Patient safety requires maintaining human oversight for high-impact AI decisions, particularly when externally matched comparator arms reduce participant exposure to standard care or when AI-derived surrogate end points replace hard clinical outcomes, which has been highlighted by the FDA–EMA joint guiding principles. AI-enabled trials also generate substantially larger and more diverse data streams, amplifying cybersecurity and privacy risks governed by the General Data Protection Regulation in Europe and the Health Insurance Portability and Accountability Act in the USA. Finally, the layering of multiple regulatory frameworks — including but not limited to the EU AI Act, the Medical Device Regulation, the Clinical Trials Regulation and data protection laws — creates a complex compliance landscape, making early engagement with regulatory bodies strongly advisable¹¹².

More generally, major regulators are opening clear pathways for responsible use of AI to enhance medicine. The FDA has outlined expectations for AI/machine learning across drug development, emphasizing risk-based validation, transparency of model life cycle and fit-for-purpose use alongside existing programmes for Complex Innovative Designs, Digital Health Technologies in clinical investigations and Real-World Evidence. The EMA’s reflection work similarly encourages human oversight, data quality/representativeness and governance across the AI life cycle, aligning with the qualification route for digital biomarkers. Convergence is emerging internationally (for example, Good Machine Learning Practice principles) and, taken together, these positions support AI for protocol design, enrichment, external comparators, adaptive monitoring and surrogate-based projections.

Practical recommendations

Several practical principles follow for researchers seeking to conduct AI-enabled trials. Regulators should be engaged early, through the qualification-of-novel-methodologies pathways described above, so that evidentiary expectations are agreed before a tool is embedded in a pivotal design. The intended context of use should be pre-specified (what the tool predicts, for which population and at what timepoint) as this defines the standard of evidence required. Validation should be demonstrated at both the individual level and the trial level. Results should be reported to AI-specific standards, using the CONSORT-AI and SPIRIT-AI extensions together with a standardized model description (a ‘model card’) and clear version tracking, to ensure transparency and reproducibility. Finally, meaningful human oversight should be preserved for high-impact decisions, with continuous monitoring for performance drift once a tool is deployed.

Outlook

The concept we propose, an AI-enabled clinical trial framework, is based on carefully selected, fit-for-purpose AI tools that enable clinical trials at different stages, from design and patient selection to data monitoring and outcome prediction. It could generalize to any field where long-term end points, heterogeneous trajectories and cost impede progress⁴⁶. However, specialties are at different stages of maturity; for instance, transplantation and oncology have more advanced regulatory-qualified surrogates, cardiology has pioneered mechanistic digital twins with clinical translation and neurology is earlier in the validation pathway for most AI applications. The framework is therefore aspirational across fields but not yet uniformly established.

Early experience in transplantation, using the EMA-qualified iBox surrogate, shows that development timelines (from patient enrolment to an interpretable efficacy readout) may be shortened without compromising rigour^50,113. This could become practical in neurology, oncology or cardiology as domain-specific surrogates and shared data infrastructures reach regulatory grade. In the near term, trial teams could conduct pre-trial design sprints that simulate accrual, power and end point behaviour under multiple scenarios, to operate surrogate-anchored adaptive platforms that discontinue futile arms and amplify promising signals, and to incorporate externally matched comparators that improve efficiency and ethical balance when standard control arms are problematic. Mechanistic maps that connect biomarkers, interventions and outcomes could guide end point selection and patient stratification, while transparent reporting with CONSORT and CONSORT-AI appendices¹¹⁴, model cards, versioning and bias or uncertainty audits could raise reproducibility and interpretability.

The pace of research in this area is accelerating along several converging fronts: increasingly autonomous LLM agents for automated clinical trial outcome prediction¹¹⁵; digital twin ecosystems for oncology trial operations including LLM-based pan-cancer digital twins¹¹⁶ and multimodal cognitive digital twins for predicting decline in Alzheimer’s disease¹¹⁷; and Bayesian extensions of prognostic covariate adjustment methods that further increase trial power while aligning with the FDA’s risk-based credibility framework for AI³⁴. These innovations will naturally integrate into the AI-enabled clinical trial framework we propose, provided each advance is matched by fit-for-purpose validation and human oversight.

The objective is not to replace human expertise but to expand foresight, compress development cycles and raise evidentiary standards in a collaborative AI–human loop. With disciplined data stewardship and close regulatory partnership, enabling clinical trials with AI could become the default route to faster, fairer and more reliable therapeutic evaluation across transplantation, neurology, oncology, cardiology and beyond.

Supplementary Material

Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s44222-026-00487-7.

Key points.

• Randomized controlled trials remain the evidentiary gold standard but are slowed by cost, complex eligibility, recruitment, long follow-up for hard end points and literature noise from unreliable studies.

• We propose an artificial intelligence (AI)-enabled clinical trial engineering framework whose ambition is to improve the design and conduct of clinical trials.

• Deeply phenotyped cohorts can provide externally matched comparator arms, and agentic AI can assist in data collection and curation.

• Regulatory-qualified surrogates (such as iBox in kidney transplantation or annualized relapse rate in neurology) can shorten trials while preserving rigorous benefit–risk assessment, and digital twins can simulate patient trajectories.

• All these tools can improve trial design; however, they need fit-for-purpose validation, human oversight and life cycle governance.

• Applied carefully, AI-enabled clinical trials could speed access to effective therapies and earlier elimination of futile or harmful ones.

Related links

BEN-8744 (BenevolentAI): https://www.benevolent.com/news-and-media/press-releases-and-in-media/benevolentai-announces-positive-topline-safety-and-pharmacokinetic-data-phase-ia-clinical-study-ben-8744-healthy-volunteers/

Clinical Trial Innovation program: https://www.fda.gov/about-fda/center-drug-evaluation-and-research-cder/cder-center-clinical-trial-innovation-c3ti

Clinical Trials Guidance Documents: https://www.fda.gov/science-research/clinical-trials-and-human-subject-protection/clinical-trials-guidance-documents

Clinical Trials Regulation: https://www.ema.europa.eu/en/human-regulatory-overview/research-development/clinical-trials-human-medicines/clinical-trials-regulation

Complex Innovative Designs: https://www.fda.gov/drugs/development-resources/complex-innovative-trial-design-meeting-program

Conducting Clinical Trials with Decentralized Elements: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/conducting-clinical-trials-decentralized-elements

Dassault Systèmes — MEDITWIN: https://www.3ds.com/science/meditwin

Digital Health Technologies in clinical investigations: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/digital-health-technologies-remote-data-acquisition-clinical-investigations

Draft guidance: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing

EMA qualification pathways: https://www.ema.europa.eu/en/human-regulatory-overview/research-development/scientific-advice-protocol-assistance/qualification-novel-methodologies-medicine-development

EMA qualified (iBox Scoring System): https://www.ema.europa.eu/en/documents/scientific-guideline/qualification-opinion-ibox-scoring-system-secondary-efficacy-endpoint-clinical-trials-investigating-novel-immunosuppressive-medicines-kidney-transplant-patients_en.pdf

EMA qualified: https://www.ema.europa.eu/en/documents/regulatory-procedural-guideline/qualification-opinion-prognostic-covariate-adjustment-procovatm_en.pdf

EMA’s reflection work: https://www.ema.europa.eu/en/about-us/how-we-work/data-regulation-big-data-other-sources/artificial-intelligence

ENGAGE: https://clinicaltrials.gov/study/NCT02477800

Enhancing Participation in Clinical Trials — Eligibility Criteria, Enrollment Practices, and Trial Designs: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/enhancing-participation-clinical-trials-eligibility-criteria-enrollment-practices-and-trial-designs

Estimands and Sensitivity Analysis: https://www.ema.europa.eu/en/documents/scientific-guideline/ich-e9-r1-addendum-estimands-and-sensitivity-analysis-clinical-trials-guideline-statistical-principles-clinical-trials-step-5_en.pdf

EU AI Act: https://artificialintelligenceact.eu/

Externally matched comparators: https://www.fda.gov/media/164960/download

FDA credibility framework: https://www.fda.gov/media/184830/download

FDA Investigational Device Exemption: https://www.fda.gov/medical-devices/premarket-submissions-selecting-and-preparing-correct-submission/investigational-device-exemption-ide

Fenebrutinib: https://clinicaltrials.gov/study/NCT04586010

First AI tool qualified by EMA: https://www.ema.europa.eu/en/news/ema-qualifies-first-artificial-intelligence-tool-diagnose-inflammatory-liver-disease-mash-biopsy-samples

First AI-powered tool qualified by FDA: https://www.fda.gov/drugs/drug-safety-and-availability/fda-qualifies-first-ai-drug-development-tool-will-be-used-mash-clinical-trials

First autonomous AI diagnostic system: https://www.aao.org/education/headline/first-ai-screen-diabetic-retinopathy-approved-by-f

General Considerations for Clinical Studies: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/e8r1-general-considerations-clinical-studies

Good Clinical Practice: https://database.ich.org/sites/default/files/ICH_E6%28R3%29_Step4_FinalGuideline_2025_0106.pdf

Good Machine Learning Practice principles: https://www.fda.gov/medical-devices/software-medical-device-samd/good-machine-learning-practice-medical-device-development-guiding-principles

Guidance: https://www.unlearn.ai/blog/us-fda-comments-on-unlearns-procova-methodology

Guiding Principles: https://www.fda.gov/about-fda/artificial-intelligence-drug-development/guiding-principles-good-ai-practice-drug-development

NCT03536052: https://clinicaltrials.gov/study/NCT03536052

NCT03887455: https://www.fda.gov/media/169337/download

Orphan Drug Designation: https://www.globenewswire.com/news-release/2023/2/8/2604040/31533/en/Insilico-Medicine-Receives-FDA-Orphan-Drug-Designation-for-Generative-AI-Discovered-and-Designed-Drug-for-Idiopathic-Pulmonary-Fibrosis.html

Program: https://www.fda.gov/drugs/development-approval-process-drugs/drug-development-tool-ddt-qualification-programs

Real-World Evidence: https://www.fda.gov/science-research/science-and-research-special-topics/real-world-evidence

Reflection paper: https://www.ema.europa.eu/en/documents/scientific-guideline/reflection-paper-use-artificial-intelligence-ai-medicinal-product-lifecycle_en.pdf

Sandbox: https://medregs.blog.gov.uk/2025/10/31/how-the-ai-airlock-is-charting-a-path-for-regulating-ai-in-healthcare-through-sandboxes/

Statistical Principles: https://www.ema.europa.eu/en/ich-e9-statistical-principles-clinical-trials-scientific-guideline

Surrogate end points: https://www.fda.gov/drugs/development-resources/table-surrogate-endpoints-were-basis-drug-approval-or-licensure

The basis for regulatory approval: https://www.fda.gov/drugs/development-resources/surrogate-endpoint-resources-drug-and-biologic-development