How AI Is Transforming Clinical Trials in 2026

Two regulatory milestones in early 2025 defined the current moment for AI in clinical research. In January 2025, the FDA published its inaugural draft guidance on artificial intelligence in drug development, titled "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products" (Docket FDA-2024-D-4689) ^[1]. That same month, on January 6, 2025, ICH E6(R3), the first comprehensive overhaul of Good Clinical Practice guidance since 2016, reached Step 4 adoption, establishing a risk-proportionate, technology-neutral framework that explicitly accommodates electronic health record integrations, wearables, and decentralized trial models ^[2].

The FDA did not issue that guidance because AI was arriving. It issued it because AI was already embedded across drug development, and the regulatory architecture was catching up. According to an FDA press announcement issued alongside the guidance, the agency had accumulated experience with more than 500 drug and biological product submissions containing AI components since 2016 ^[3].

The operational context for this shift is well-documented. As of early 2025, ClinicalTrials.gov had surpassed 530,000 registered studies ^[16]. Actual enrollment timelines have historically run nearly double planned timelines, according to Tufts Center for the Study of Drug Development (Tufts CSDD) research covering more than 150 Phase II and III studies across nearly 16,000 investigative sites ^[5]. The total cost of developing a single approved drug, including the cost of failures, was estimated at $2.6 billion (2013 dollars) in the most frequently cited Tufts CSDD analysis by DiMasi and colleagues ^[6]. Against that scale of expenditure and delay, the question is not whether AI has a role in clinical trials. The question is where it is producing verified, rather than anticipated, results, and what regulatory expectations and legal obligations are relevant to those deployments.

This article examines both.

Why the Evidence Base Is Now More Than Theoretical

For several years, AI in clinical trials was largely discussed as potential. The published literature before 2020 was dominated by feasibility demonstrations and proof-of-concept models. That has changed. A 2025 comprehensive narrative review covering PubMed, Embase, IEEE Xplore, and Google Scholar databases from January 2015 to December 2024 analyzed a substantial body of AI applications across the clinical trial lifecycle, including patient recruitment, safety monitoring, site selection, and outcome prediction ^[4]. A separate 2025 scoping review analyzed 142 published studies on AI in clinical trial risk assessment, covering safety, efficacy, and operational risk prediction ^[7]. These are not isolated case studies; they represent a literature that, while heterogeneous in methodology and quality, now spans over a decade of deployment experience.

What the literature does not yet offer consistently is prospective, controlled validation of AI tools at scale across multiple therapeutic areas and geographies. Many published studies are retrospective, simulated against historical data, or conducted within single institutions. Readers should apply that limitation when evaluating any performance figures described in this article.

The FDA and the Clinical Trial Transformation Initiative (CTTI), a public-private partnership co-founded by Duke University and the FDA that collaborates with more than 500 organizations spanning academia, industry, regulators, and patient advocates, convened a public workshop in August 2024 to discuss guiding principles for responsible AI adoption across drug and biological product development ^[8]. The workshop report identified four primary hurdles: limited access to relevant data, low confidence in AI model outputs, siloed disciplines that slowed cross-functional implementation, and regulatory complexity surrounding AI-specific requirements.

Where AI Is Operating Across the Trial Lifecycle

Protocol Design and Feasibility

Protocol amendments are among the most disruptive and costly mid-trial events a sponsor can face. Tufts CSDD research has estimated per-amendment costs at significant levels, with complexity and phase strongly influencing total impact ^[9]. Published Tufts analyses have identified eligibility criteria among the leading drivers of unplanned amendments, alongside changes in sponsor strategy and new safety information ^[9]. AI approaches this problem from two directions. Natural language processing tools can analyze historical protocol structures, identify eligibility criteria patterns associated with enrollment failures in prior trials, and flag internal inconsistencies between the protocol synopsis and body. Separately, EHR-linked feasibility models can simulate patient availability against proposed inclusion and exclusion criteria before a trial launches.

A 2023 paper published in Communications Medicine described research using AI-based tools, including an open-source model called Trial Pathfinder, which used EHR data to simulate completed non-small-cell lung cancer trials under varying eligibility criteria ^[10]. The analysis found that relaxing selected overly restrictive criteria approximately doubled the simulated eligible patient pool without altering the direction of the trial's survival findings. This is a promising retrospective simulation result in one indication; it does not establish that implementing AI-informed eligibility modification prospectively will produce the same outcome across indication types or geographic settings.

Generative AI models are also being applied to protocol drafting: producing structured sections, checking internal consistency, and flagging discrepancies between time-and-events tables and in-text procedure descriptions. Based on the literature reviewed for this article, peer-reviewed validation of these drafting capabilities in published clinical trial settings remains limited; most available evidence comes from developer demonstrations and early adopter reports rather than controlled studies. These tools do not replace scientific and medical judgment. They reduce time spent on mechanical consistency review, allowing clinical teams to focus on the decisions that depend on expertise.

Site Selection and Enrollment Forecasting

Site selection has traditionally depended on investigator relationships, incomplete historical performance records, and feasibility questionnaires completed under time pressure. The predictable result: sites that look strong on paper underperform, and enrollment targets slip. Tufts CSDD research found that actual enrollment timelines across more than 150 Phase II and III studies ran nearly double their planned timelines, and that even when trials ultimately met their goals, the timeline extensions were substantial ^[5]. Approximately 37% of investigative sites under-enroll in a given study, and about 11% fail to enroll a single patient ^[5].

Machine learning models trained on historical site-level data offer a more systematic approach. A 2024 study published in PLOS ONE by researchers from Janssen Pharmaceuticals developed a machine learning model using historical recruitment records and real-world data to rank candidate sites by predicted enrollment performance ^[11]. The model outperformed conventional industry baselines in the two indications tested, drawing on site demographics, local disease prevalence, and prior site performance. This is encouraging evidence, though the authors appropriately note that the model's generalizability across other indications requires further validation.

Predictive analytics platforms trained on large multi-sponsor datasets can now evaluate protocol characteristics, regional epidemiology, investigator track records, and patient demographics simultaneously, generating ranked site recommendations early enough in trial planning to shape site activation strategy ^[4]. One operationally important application: these models can flag geographic and demographic coverage gaps before enrollment opens, supporting FDA's recommendations for representative trial populations. FDA's December 2025 final guidance "Enhancing Participation in Clinical Trials" recommends that sponsors broaden eligibility criteria, optimize trial sites, and monitor enrollment to increase representation across demographic and non-demographic characteristics ^[17]. AI-assisted site selection can support that process upstream, though the guidance addresses representative enrollment generally rather than validating any specific AI tool.

Patient Recruitment and Eligibility Screening

Patient recruitment is where the distance between conventional methods and AI-assisted approaches is most visible in operational terms. Enrollment timeline slippage is persistent across the industry. In the 2013 Tufts CSDD analysis, timelines ran nearly double planned lengths even when trials ultimately reached enrollment targets; a 2020 Tufts CSDD follow-up study covering 87 trials found meaningful improvement, with 77% of studies having enrollment timelines equal to or shorter than planned, up from 48% in 2012 ^[12]. Progress has been made, but enrollment remains among the most unpredictable phases of clinical development.

FHIR-connected EHR screening tools can apply a trial's eligibility criteria programmatically against structured patient records, surfacing candidates who would be missed by manual chart review. A 2023 paper in Communications Medicine reviewed earlier AI-based patient-trial matching systems in oncology and reported generally high accuracy in identifying eligible candidates from EHR data, while noting that comparability across systems is difficult without a defined gold standard for eligibility assessment ^[10]. These evaluations are predominantly retrospective; results depend heavily on EHR data completeness, institution type, and the precision with which eligibility criteria were encoded.

The selection bias inherent in manual recruitment is worth naming directly. Sites with more staff and stronger investigator networks tend to identify more candidates. Sites operating with fewer resources screen fewer records. AI-assisted screening applies consistent criteria regardless of site staffing, which can reduce this particular form of variability. That said, if the underlying eligibility criteria or EHR data themselves encode demographic or geographic bias, automated screening will replicate rather than correct it.

Data Management and Safety Monitoring

Clinical trial data volumes have grown alongside protocol complexity. Remote and decentralized trial elements add continuous data streams from wearables, patient-reported outcome applications, and sensor-based devices, all requiring timely interpretation.

AI and machine learning methods are being applied to pharmacovigilance and adverse event detection with increasing sophistication. A 2025 study published in ScienceDirect proposed a hybrid AI framework integrating structured patient data with unstructured clinical notes, using deep learning and NLP models to detect serious adverse events ^[13]. The investigators report that the framework outperformed conventional signal-detection methods in testing on its study dataset; this result is study-specific and has not been independently replicated, and performance in different clinical or geographic contexts has not been established.

The 2025 comprehensive literature review covering 2015-2024 studies examined AI applications in continuous monitoring settings across heterogeneous populations, endpoints, and comparator methods ^[4]. Any performance figures from that synthesis reflect the range of individual studies reviewed rather than a single validated benchmark.

NLP systems designed for adverse event extraction from clinical notes, including systems like MADEx (Medication and Adverse Drug Event Extraction), can process large volumes of clinical documentation more rapidly than conventional periodic manual review workflows ^[14]. Earlier signal identification may enable faster escalation to safety monitoring boards. Whether this throughput advantage translates into fewer missed events in actual trial operations depends on implementation factors, alert fatigue thresholds, and site workflow integration that the extraction studies do not address.

Regulatory Documentation

A single Phase III program generates thousands of pages spanning protocol, investigator's brochure, informed consent forms, development safety update reports, statistical analysis plan, and clinical study report. Each document must be internally consistent, cross-referenced correctly, and compliant with applicable requirements including ICH E6(R3) ^[2], ICH E3 ^[18], ICH E2F ^[19], and jurisdiction-specific formatting standards.

Generative AI tools designed for regulatory document production are designed to reduce the time required to produce first drafts of structured documents. They also offer cross-document consistency checking that can catch discrepancies a manual reviewer may miss, particularly in large multi-protocol programs where protocol amendments cascade through the ICF, site training materials, and regulatory correspondence.

The separation between AI used for regulated evidence generation and AI used for administrative productivity matters here. FDA guidance (FDA-2024-D-4689) applies to AI where the output is intended to support regulatory decisions about safety, effectiveness, or quality ^[1]. AI used purely for drafting efficiency, with qualified human reviewers taking full responsibility for the submitted text, falls into a different operational category. Organizations deploying AI in regulatory writing should conduct a clear scoping exercise to determine whether the tool's output is entering the regulatory evidence chain or the administrative workflow.

The Regulatory Framework: What FDA and ICH Now Recommend

The FDA's January 2025 draft guidance (FDA-2024-D-4689) is the most specific statement the agency has issued on AI in drug development ^[1]. It introduces a risk-based credibility assessment framework built around "context of use" (COU), defined as the specific role and scope of an AI model in addressing a regulatory question. Establishing and documenting credibility involves a seven-step process: defining the question of interest, characterizing the COU, assessing model risk, planning model evaluation, executing evaluation activities, documenting and communicating results, and maintaining the model over its lifecycle.

Two aspects of this framework deserve careful attention. First, it applies specifically to AI models where the output is intended to support regulatory decisions about safety, effectiveness, or quality. AI used for internal operational efficiency that does not influence regulatory submissions is outside this guidance's direct scope. Second, the guidance is currently a draft and provides recommendations rather than binding requirements. Sponsors should monitor the FDA docket for finalization and treat the current framework as the agency's stated direction.

The guidance addresses model transparency and the importance of being able to explain how an AI model reaches its conclusions, particularly where the COU carries higher risk. This does not categorically prohibit complex models, but it means sponsors should document the reasoning behind model design choices, validation approaches, and limitations relevant to the regulatory context. The guidance's risk-proportionate framework is consistent with the good-practice expectation that qualified experts evaluate AI outputs informing regulatory submissions, though this is an inference from the framework's principles rather than an explicit universal rule. Documentation expectations around that review may be clarified if and when the guidance is finalized; final FDA guidances typically remain non-binding and clarify recommendations rather than establish mandatory requirements ^[1].

ICH E6(R3), adopted at Step 4 on January 6, 2025, establishes a risk-proportionate quality management framework for interventional clinical trials ^[2]. It is technology-neutral: it accommodates wearables, eConsent, EHR integrations, and decentralized models, provided that data integrity, audit trails, and participant protection are maintained. For AI-assisted monitoring specifically, E6(R3) establishes quality system, validation, and traceability principles relevant to centralized data review. It is not an AI-specific regulation, but its principles of quality by design and risk-proportionate oversight apply directly to organizations building AI-assisted workflows into their trial operations.

Limitations and Where Caution Is Warranted

No account of AI in clinical trials is complete without its constraints. Several deserve explicit attention.

Training data representativeness. AI models perform within the boundaries of their training populations. Systems trained predominantly on data from academic medical centers in North America or Western Europe may not generalize reliably to sites in different demographic, geographic, or healthcare infrastructure contexts. Selection bias in training data propagates into model outputs and predictions.

Evidence maturity. Most published AI validation studies in clinical research are retrospective, simulated, or conducted within single institutions. Prospective, multi-site validation across diverse indications and populations remains limited. Sponsors evaluating commercial AI platforms should ask specifically whether published performance evidence is retrospective or prospective, and whether the validation population matches their target context.

False positives, alert fatigue, and workflow burden. AI safety monitoring systems that flag adverse events with high sensitivity may also generate false positive signals. Sites managing alert fatigue can inadvertently deprioritize genuine signals. Deployment decisions should account for the full workflow impact, not only detection sensitivity figures from validation studies.

Model drift. AI models should be monitored for performance degradation as trial populations, site practices, and data entry behaviors evolve. A model validated at trial start may perform differently at 18 months if input characteristics shift. Including model performance monitoring in the maintenance plan is consistent with FDA's recommended lifecycle approach in the draft guidance, though the specific form and frequency of monitoring are not prescribed in the current non-binding draft ^[1].

EU regulatory alignment. The EU AI Act, which entered into force in August 2024, establishes a risk-tiered classification system for AI across sectors, including healthcare ^[15]. AI systems used in clinical settings do not automatically fall into a specific risk category; classification depends on the system's intended purpose, the specific provision applicable to that purpose, and a legal analysis of the system's function. Organizations running trials across EU member states should conduct an AI Act classification assessment for each AI system in their clinical operations before deployment, rather than assuming classification from the healthcare sector alone. The phased schedule for obligations under the EU AI Act means that requirements continue to come into force on different timelines through 2026 and beyond.

Explainability versus accuracy trade-offs. More complex models are sometimes harder to explain to regulators and may require more extensive credibility evidence under FDA's risk-based framework, depending on the COU. The relationship between model complexity and performance is use-case dependent. Organizations should evaluate explainability requirements against the COU's risk level, rather than assuming either that the highest-performing model will pass regulatory scrutiny or that the most explainable model will necessarily perform adequately.

How Kitsa Fits Into This Problem

Kitsa describes itself as an AI-native clinical research infrastructure platform. According to the company's product documentation, KScribe is designed to generate regulatory documents including protocols, ICFs, IBs, DSURs, and CSRs, with cross-document consistency checking built into the workflow. KScout is designed to apply predictive analytics to site feasibility and enrollment forecasting. KScreener is designed to use FHIR-connected EHR queries to surface candidate patients against eligibility criteria more systematically than manual chart review. Kitsa states that the platform is SOC 2 Type II attested, HIPAA compliant, and ISO 27001 certified, deployed in a private AWS VPC environment. These are vendor-stated specifications; independent performance validation data is not publicly available.

Key Takeaways

• The FDA published its first AI-specific draft guidance for drug and biological products in January 2025 (Docket FDA-2024-D-4689), introducing a risk-based, context-of-use credibility assessment framework. The guidance provides recommendations and is not yet binding pending finalization.
• ICH E6(R3) reached Step 4 adoption on January 6, 2025, establishing a technology-neutral, risk-proportionate GCP framework that accommodates EHR integrations, wearables, and decentralized trial models, with EU implementation of Principles and Annex 1 effective from July 2025.
• Enrollment timeline slippage has historically been a persistent industry challenge. Tufts CSDD data covering more than 150 Phase II and III studies found actual timelines ran nearly double planned timelines; a 2020 follow-up showed 77% of trials with enrollment timelines equal to or shorter than planned, indicating improvement but not full resolution.
• ML models for site selection trained on historical enrollment records and real-world data have demonstrated improved performance over conventional baselines in peer-reviewed validation studies, though generalizability across indications requires further prospective evidence.
• Most published AI performance figures in clinical research come from retrospective or single-institution studies. Sponsors evaluating AI tools should specifically ask whether validation evidence is prospective, multi-site, and matched to their target indication and geography.
• The FDA's risk-proportionate credibility framework is consistent with the good-practice expectation that qualified experts evaluate AI outputs informing regulatory submissions. This is an inference from the framework's principles; documentation expectations may be clarified if and when the guidance is finalized as non-binding recommendations.
• The EU AI Act does not automatically classify clinical trial AI systems as high-risk; classification requires intended-purpose analysis under the applicable provisions of the legislation.

FAQ

What is the FDA's current position on AI use in clinical trials?

+

In January 2025, the FDA published a draft guidance (Docket FDA-2024-D-4689) that recommends sponsors applying AI to support regulatory decisions establish model credibility through a seven-step, risk-based framework centered on each model's context of use. When finalized, the guidance will provide recommendations concerning AI models used across the nonclinical, clinical, post-marketing, and manufacturing phases of drug and biological product development, where the output is intended to support regulatory decisions. Final FDA guidance documents typically remain non-binding ^[1].

Can AI replace manual patient screening in clinical trials?

+

AI-assisted screening tools can apply eligibility criteria systematically across large EHR databases and identify candidates faster than manual chart review. Published studies in oncology have demonstrated generally high accuracy in retrospective evaluations, though comparability across systems is limited without a standard gold standard for eligibility assessment ^[10]. Final enrollment decisions require review by the investigator or appropriately qualified delegated trial personnel. AI tools change the ratio of records requiring human attention, not the clinical judgment required to enroll a patient.

What types of AI are being used in clinical trials today?

+

Current applications span NLP for eligibility screening and regulatory document drafting, machine learning for site selection and enrollment forecasting, deep learning for adverse event detection, and predictive analytics for adaptive design and outcome prediction. A 2025 scoping review of 142 published studies documented AI applications across safety, efficacy, and operational risk prediction, with large language models appearing in 7 of 33 studies published in 2023 alone ^[7].

What are the main risks of using AI in clinical trials?

+

Key risks include training data bias toward particular populations or healthcare systems, false positive signals that create alert fatigue in safety monitoring, performance drift as trial populations and data entry practices evolve, and explainability limitations in complex models that face regulatory scrutiny under FDA's credibility framework. Organizations should also assess change control implications when updating or replacing AI systems mid-trial ^[1].

How does AI help with regulatory documentation in clinical trials?

+

Generative AI tools can draft structured regulatory documents and check cross-document consistency, such as flagging when a protocol amendment has not been reflected in the informed consent form or site training materials. These tools are particularly relevant in large multi-document programs where manual consistency review is resource-intensive. AI-assisted documents should undergo sponsor-defined review and approval by appropriately qualified personnel before submission; AI drafting does not transfer responsibility for submission accuracy.

Does the EU AI Act affect clinical trial AI systems?

+

Yes, but the classification implications require careful analysis. The EU AI Act, which entered into force in August 2024, applies a risk-tiered framework to AI systems across sectors ^[15]. Whether a specific clinical trial AI application falls into a high-risk category depends on its intended purpose and which provision of the legislation applies, not on the healthcare sector alone. Organizations should conduct a classification assessment for each system and monitor the Act's phased implementation schedule, as different obligations continue entering into force through 2026 and beyond.

References

1. [1]U.S. Food and Drug Administration. "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products." Draft Guidance for Industry, Docket FDA-2024-D-4689. January 7, 2025. https://www.federalregister.gov/documents/2025/01/07/2024-31542/considerations-for-the-use-of-artificial-intelligence-to-support-regulatory-decision-making-for-drug
2. [2]International Council for Harmonisation. "ICH E6(R3): Guideline for Good Clinical Practice." Step 4 Final Version, adopted January 6, 2025. https://database.ich.org/sites/default/files/ICH_E6%28R3%29_Step4_FinalGuideline_2025_0106.pdf
3. [3]U.S. Food and Drug Administration. "FDA Proposes Framework to Advance Credibility of AI Models Used for Drug and Biological Product Submissions." FDA Press Announcement, January 6, 2025. Reports FDA experience with more than 500 drug and biological product submissions containing AI components since 2016. https://www.fda.gov/news-events/press-announcements/fda-proposes-framework-advance-credibility-ai-models-used-drug-and-biological-product-submissions
4. [4]Almutairi SM, et al. "Artificial Intelligence in Clinical Trials: A Comprehensive Review of Opportunities, Challenges, and Future Directions." International Journal of Medical Informatics, ScienceDirect, 2025. https://www.sciencedirect.com/science/article/pii/S1386505625003582
5. [5]Tufts Center for the Study of Drug Development. "New Research Characterizes Effectiveness and Variability of Patient Recruitment and Retention Practices." January 15, 2013. Analyzed 150+ clinical studies, approximately 16,000 investigative sites, Phase II and III trials. https://www.fiercebiotech.com/biotech/new-research-from-tufts-center-for-study-of-drug-development-characterizes-effectiveness
6. [6]DiMasi JA, Grabowski HG, Hansen RW. "Innovation in the Pharmaceutical Industry: New Estimates of R&D Costs." Journal of Health Economics, 2016;47:20-33. DOI: 10.1016/j.jhealeco.2016.01.012. Total estimated cost of $2.6 billion per approved drug (2013 dollars), including cost of capital and failures. https://doi.org/10.1016/j.jhealeco.2016.01.012
7. [7]Teodoro D, et al. "A Scoping Review of Artificial Intelligence Applications in Clinical Trial Risk Assessment." NPJ Digital Medicine, July 30, 2025. Analyzed 142 published studies (2013-2024). https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12307910/
8. [8]U.S. Food and Drug Administration and Clinical Trial Transformation Initiative. "Leveraging Artificial Intelligence in Drug and Biological Product Development: FDA-CTTI Public Workshop Report." PubMed Central, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12690500/
9. [9]Getz KA, et al. "Measuring the Incidence, Causes, and Repercussions of Protocol Amendments." Drug Information Journal, 2011;45(3):265-275. DOI: 10.1177/009286151104500308. Referenced in Tufts CSDD analyses of protocol amendment costs. https://doi.org/10.1177/009286151104500308
10. [10]Zhang B, et al. "Harnessing Artificial Intelligence to Improve Clinical Trial Design." Communications Medicine, December 21, 2023. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10739942/
11. [11]Hulstaert L, et al. "Enhancing Site Selection Strategies in Clinical Trial Recruitment Using Real-World Data Modeling." PLOS ONE, March 11, 2024. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10927105/
12. [12]Tufts Center for the Study of Drug Development. "Progress Seen in Streamlining of Patient Recruitment and Retention for Trials." Reported by ACRP, January 29, 2020. Follow-up study (87 trials): 77% had enrollment timelines equal to or shorter than planned, up from 48% in 2012 baseline. https://acrpnet.org/2020/01/29/progress-seen-in-streamlining-of-patient-recruitment-and-retention-for-trials
13. [13]Al-Rashidi A, et al. "AI-Driven Pharmacovigilance: Enhancing Adverse Drug Reaction Detection with Deep Learning and NLP." ScienceDirect, June 2025. https://www.sciencedirect.com/science/article/pii/S221501612500305X
14. [14]Biran O, et al. "Artificial Intelligence in Pharmacovigilance: Advancing Drug Safety Monitoring and Regulatory Integration." PubMed Central, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12317250/
15. [15]European Parliament and Council of the European Union. "Regulation (EU) 2024/1689: Artificial Intelligence Act." Official Journal of the European Union, August 2024. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
16. [16]National Library of Medicine. "ClinicalTrials.gov: A 25-Year Journey to a Half-Million Registered Studies." NLM Director's Blog, April 2, 2025. More than 530,000 total studies as of April 2025. https://nlmdirector.nlm.nih.gov/2025/04/02/clinicaltrials-gov-a-25-year-journey-to-a-half-million-registered-studies/
17. [17]U.S. Food and Drug Administration. "Enhancing Participation in Clinical Trials: Eligibility Criteria, Enrollment Practices, and Trial Designs." Final Guidance for Industry, December 15, 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/enhancing-participation-clinical-trials-eligibility-criteria-enrollment-practices-and-trial-designs
18. [18]International Council for Harmonisation. "ICH E3: Structure and Content of Clinical Study Reports." Guideline, November 1995. https://www.ich.org/page/efficacy-guidelines
19. [19]International Council for Harmonisation. "ICH E2F: Development Safety Update Report." Guideline, August 2011. https://www.ich.org/page/safety-guidelines

Suggested internal links: KScribe · KScout · KScreener