Clinical trial protocols have grown substantially more complex over the past two decades. Research drawing on sponsor and CSDD data found that the average number of protocol-specified procedures rose by roughly 65% between 2000 and 2020, driven by innovations in precision medicine, biomarker-stratified designs, and the integration of patient-reported outcomes.³ This complexity creates more points of failure: eligibility criteria are set before confirmatory enrolment data exist; statistical assumptions about dropout and baseline variability may not hold at scale; and regulatory requirements evolve between protocol approval and first patient enrolment. The result is a high and persistent rate of post-submission amendments that impose remediation costs well more than their technical complexity.

The CRF is the primary instrument through which protocol requirements are operationalised into structured data. Its quality directly determines downstream data integrity: a field with an ambiguous response format, an unconstrained free-text entry where coded terminology is required, or a mandatory field whose clinical rationale was never made explicit to site staff will generate queries at a rate that erodes both the efficiency of data management and the quality of the final dataset submitted to regulators.⁴ Mapping CRF data to Clinical Data Interchange Standards Consortium (CDISC) Study Data Tabulation Model (SDTM) standards for regulatory submission adds a further layer of technical work that, when done manually, is time-consuming and error-prone. It is important to distinguish the two CDISC standards referenced throughout this review: CDASH (Clinical Data Acquisition Standards Harmonization) standardises how data are structured at the point of collection—that is, CRF field design—while SDTM standardises how that same data is organised for regulatory submission. AI-assisted mapping tools discussed in this review typically operate at the CDASH-to-SDTM transformation step, translating collection-ready fields into submission-ready domains; the two standards serve sequential, not interchangeable, purposes.⁵ Errors in SDTM mapping can result in submission deficiencies that delay marketing authorisation decisions—a downstream consequence rarely traced back to the CRF design decision that caused it.⁵

A note on burden. Throughout this review, reductions in CRF field count or query volume are described as reducing site and data management burden, i.e., the workload associated with data entry, source-document verification, query resolution, and database management. This is distinct from patient burden, which is determined by patient-facing procedures such as visits, blood draws, imaging, and questionnaires. Removing an unnecessary CRF variable reduces the work required to record and clean a data point; it does not, by itself, reduce what is asked of the trial participant unless the underlying procedure is also removed. Where a cited source specifically reports a reduction in patient-facing procedures, this review states so explicitly; where a source reports only CRF-field or query-rate reduction, this review characterises the finding as a site/data-management burden reduction, not a patient burden reduction.

Because this review synthesises material of markedly different maturity—peer-reviewed studies, regulatory qualification opinions, proof-of-concept reports, vendor white papers, industry blogs, and single-sponsor observational implementations—sources are classified into five evidence tiers, applied consistently in Tables 1, 2, and 4 and cited alongside each quantitative claim in the text:

No formal quality-assessment instrument (e.g., GRADE, Newcastle-Ottawa) was applied as the heterogeneity of source types (regulatory documents, vendor reports, peer-reviewed studies) makes a single instrument poorly suited to this body of evidence; the five-tier classification above serves as a transparent, source-type-based proxy for evidence strength, and readers requiring formal quality appraisal of an individual Tier A/B study are directed to the primary source.

Rather than describing GenAI and predictive analytics as a single undifferentiated category, this review organises the technologies discussed into four functional groups, each with distinct objectives, validation strategies, and regulatory implications. This taxonomy is used consistently in the Results section below.

These four categories differ in an important respect for readers evaluating the evidence: Document Generation and Metadata Automation tools produce artefacts intended for mandatory human expert review before use, whereas Operational Prediction and Statistical Augmentation outputs can, in some configurations, feed more directly into statistical or operational decisions, which is also why they attract closer regulatory and biostatistical scrutiny, as discussed below.

This narrative review drew on peer-reviewed literature retrieved from PubMed, Embase, Scopus, and IEEE Xplore, with a search date range of January 2019 through March 2025. Search terms included combinations of: “generative AI”, “large language model”, “GPT”, “predictive analytics”, “machine learning”, “digital twin”, “clinical trial protocol”, “case report form”, “CRF design”, “CDISC”, “SDTM”, “protocol amendment”, “electronic data capture”, and “adaptive trial design”. Results were supplemented by the US Food and Drug Administration (FDA) and European Medicines Agency (EMA) regulatory guidance documents, International Council on Harmonisation (ICH) guidelines (specifically E6(R3) and E8(R1)), and grey literature from pharmaceutical sponsors and contract research organisations where methodological detail was sufficient for evaluation.

Records retrieved across the four databases were merged and de-duplicated using reference-manager software prior to screening. Screening proceeded in two stages: (1) title/abstract screening for topical relevance, applying the inclusion criterion that a source directly address AI or predictive-analytics applications in protocol development, CRF construction, or CDISC compliance; and (2) full-text assessment against a second criterion: sufficient methodological transparency (a stated comparator, an identifiable data source, and a reported outcome metric) to permit critical evaluation. Sources failing either criterion, including duplicate or superseded reports, were excluded. Of approximately 190 candidate sources initially identified, 89 were retained (Figure 1).

Screening and eligibility assessment were performed by a single reviewer (the author). This is a material limitation: no second independent screener was used, formal inter-rater agreement statistics (e.g., Cohen’s kappa) are not available, and exact record counts at each intermediate screening stage were not contemporaneously logged in a form that permits retrospective reconstruction of a complete PRISMA-ScR count-by-count flow diagram. Figure 1 therefore reports only the two counts that can be stated with confidence—the initial candidate pool (n = 190) and the final included set (n = 89)—and this gap is disclosed explicitly rather than backfilled with reconstructed figures.

Grey literature (vendor white papers, platform blogs, sponsor-issued reports) was eligible for inclusion where it met the same transparency criterion; where a grey-literature source reported a quantitative claim without a stated comparator or observation period, the claim is retained here only if it is explicitly labelled Tier D or E and is not used to anchor a headline conclusion. No formal critical-appraisal checklist (e.g., AMSTAR, GRADE) was applied for the reason given in the Evidence Classification Framework above; each finding is instead tagged with its evidence tier at first mention and in Tables 1, 2, and 4.

Narrative synthesis was used rather than meta-analysis, given the heterogeneity of study designs, outcome definitions, and the predominance of proof-of-concept and observational reports in the current evidence base. Findings were organised into the four functional technology categories defined in the Background (Document Generation, Metadata Automation, Operational Prediction, Statistical Augmentation), followed by regulatory alignment and implementation challenges.

Several commercial platforms have demonstrated the feasibility of converting a structured study synopsis into an initial full protocol draft within hours rather than weeks. Systems such as Protocol AI (Risklick) and AuroraPrime RMA (AlphaLife Sciences) couple LLMs with RAG knowledge bases drawn from regulatory guidance documents, CDISC standards, and sponsor-specific historical protocols, generating section-level draft text organised around protocol-type templates encoding ICH E6(R3) and ICH E8(R1) structural requirements (Tier D: platform-reported).^8–9 A published proof-of-concept using GPT-4 with domain-specific fine-tuning demonstrated cosine similarity scores of up to 0.80 and Recall-Oriented Understudy for Gisting Evaluation (ROUGE) precision of up to 0.70 between AI-generated and expert-authored introduction sections (Tier B). These metrics indicate substantial lexical and semantic overlap with an expert reference text, though neither metric assesses clinical judgement, factual accuracy, or regulatory adequacy directly.¹⁰

On evaluation methodology beyond ROUGE and cosine similarity: both metrics quantify surface-level textual similarity to a reference document and are known to correlate imperfectly with expert judgements of factual correctness or clinical appropriateness in health-text generation tasks.²¹ Several of the sources reviewed here supplement these metrics with structured expert-rater review (e.g., completeness against a section checklist, or a binary/Likert acceptability rating from a clinical reviewer), but a standardised, publicly validated rubric for evaluating AI-generated protocol or CRF text—covering completeness, factual accuracy, and regulatory-compliance dimensions separately—does not yet exist across the sources reviewed; this is identified as a priority methodological gap (Table 4).

Amendment prevention represents a particularly valuable application within this category. By analysing patterns in historical amendment data, AI tools can flag protocol sections associated with high revision risk: eligibility thresholds inconsistent with disease epidemiology, visit schedules infeasible at typical site staffing levels, and statistical assumptions at odds with published natural history data. A synthesis from the Duke Clinical Research Institute estimated that systematic adoption of AI-based protocol simulation could reduce major amendment incidence by 20–30% across Phase II and III programmes (Tier D/E: think-tank synthesis).¹² Early sponsor-reported data suggest first-draft amendment rates declining by up to 28% in AI-assisted workflows versus historical comparators (Tier D: single-sponsor observational), though independent confirmation is lacking.⁸ For clarity, “first-draft amendment rate” as used by the cited sources denotes the proportion of protocols requiring a substantive amendment measured against the first submitted (pre-regulatory-review) version, as distinct from amendment rates measured after internal review board/ethics committee or regulatory review; sources do not consistently specify which definition they use, which limits cross-study comparability (Table 2).

AI-assisted CRF generation treats the Schedule of Activities (SoA) table as a structured data input. NLP pipelines parse the SoA and surrounding protocol text to identify data collection requirements at each visit, then map those requirements to CDASH-standard domain templates. Systems with cross-transclusion features maintain synchronisation across related documents when SoA parameters change, reducing version inconsistencies that arise when protocol and CRF documents are managed separately (Tier D).^{9, 14} Initial (draft) form-build timelines have been reported to compress from four-to-eight weeks to under 72 hours; sources are explicit that this figure describes the draft build only, with one-to-two additional weeks required for review and regulatory-compliance validation before the database is production-ready, a distinction this review preserves throughout Table 1 rather than presenting the 72-hour figure as the total time to a validated database.¹⁴

Machine learning models trained on historical CRF data and query repositories can generate edit-check recommendations ranked by predicted query frequency and clinical significance. Adaptive eCRF systems incorporating this logic report data-query rate reductions of 22–35% relative to conventionally built comparators (Tier D).^{2, 15} For SDTM mapping, NLP tools that assign CRF fields to SDTM domains by analogy with structurally similar historical fields have achieved domain-assignment accuracy exceeding 90% for standard domains, with more complex or non-standard fields flagged for human expert review (Tier B/D). Across published evaluations, SDTM mapping and annotated CRF generation time has been reported to reduce by 40–60% compared with fully manual processes.^{5, 16} As noted in the Background, these tools operationalise the CDASH-to-SDTM transformation specifically; they do not replace CDASH-level CRF design decisions, which remain a separate, upstream step.

Reductions in CRF field count achieved through this category of tool should be read, per the burden distinction introduced in the Background, as reductions in site workload, data entry, source-document verification, and data-management effort—not, without further evidence, as reductions in patient burden. None of the sources reviewed in this subsection reported a corresponding reduction in patient-facing procedures (visits, assessments, or sample collection); where such a reduction is claimed elsewhere in the literature, it should be substantiated by a change to the visit schedule or assessment battery itself, not inferred from CRF field count alone.

NLP-based protocol analysers can cross-reference proposed eligibility criteria against natural history datasets and real-world EHR populations to identify criteria that may render target populations implausibly small. The Stanford Trial Pathfinder system demonstrated retrospectively that targeted relaxation of specific haematological and medication-based exclusion criteria could double the size of eligible patient populations in some therapeutic areas without measurable change in safety profile. For enrolment forecasting, machine learning models trained on historical site- and country-level recruitment data achieved prospective accuracy rates of approximately 85%, with reported deployment reductions in recruitment delay of around 25% (Tier B).¹¹ As noted in Table 2, the precise statistical definition of “accuracy” (e.g., mean absolute percentage error against a defined forecast horizon versus a categorical on-time/delayed classification) is not uniformly reported across the underlying studies, and this heterogeneity should be considered when comparing figures across sources.

A note on scope and terminology. This subsection touches on prognostic-covariate-adjustment statistics that fall outside clinical data management’s core disciplinary boundary; the author is not a biostatistician, and the terminology below follows exactly what each cited source reports without extrapolation. Readers intending to apply these methods to trial design should consult a qualified biostatistician, particularly regarding the regulatory acceptability of a specific covariate-adjusted analysis for a given protocol.

At the frontier of AI integration, digital-twin approaches provide prognostic covariates for adaptive statistical models. The PROCOVA-MMRM framework, which incorporates AI-generated digital-twin predictions into mixed-model repeated-measures analyses, has been reported to produce power gains described by the source as equivalent to a 25–30% increase in effective sample size, while maintaining Type I error control acceptable to regulatory reviewers in the cited applications (Tier B/C).^{7, 13} Four related but distinct concepts appear across this literature, and this review treats them as non-interchangeable:

Its first application in Alzheimer’s disease trials projected sample-size reductions sufficient to shorten development timelines by 12–18 months per programme, per the cited source’s own projection rather than an independently confirmed outcome (Tier B).¹³ Readers are directed to the primary methodological references for the full model specification before applying any of these figures to a specific trial design.^{7, 13, 22}

The regulatory environment for AI in drug development has matured considerably since 2022, though several frameworks discussed here remain in draft or early-implementation form; specific effective dates and scope limitations are stated explicitly below because earlier drafts of this review understated them.

The FDA’s January 2025 draft guidance on using AI to support regulatory decision-making introduces a context-of-use (COU) credibility framework that calibrates validation requirements to the risk associated with a specific intended use.¹⁷ As of this review, the guidance remains in draft form (public comment closed April 2025; not yet finalised), and its stated scope applies specifically to AI models used to produce data supporting a regulatory decision about a drug or biological product’s safety, effectiveness, or quality—the guidance explicitly does not extend to AI used for drug discovery or for operational efficiencies such as internal workflow support or drafting a regulatory submission.²⁴ This is a materially important scope boundary for this review: most of the Document Generation and Metadata Automation tools discussed above fall outside the COU framework’s current stated scope, whereas Operational Prediction and Statistical Augmentation tools that generate evidence feeding directly into a regulatory decision are more likely to fall within it. By late 2024, the FDA had processed several hundred submissions containing AI components in some form, reflecting substantial sponsor adoption.¹⁷

The EMA’s October 2024 Reflection Paper on AI in the medicinal product lifecycle adopts a compatible risk-based orientation, with expectations organised by development phase.¹⁸ Separately, the EMA’s Committee for Medicinal Products for Human Use (CHMP) adopted its first qualification opinion for an AI methodology on 27 February 2025: AIM-NASH, an AI-based tool that assists pathologists in scoring liver biopsies for disease activity in MASH/NASH clinical trials.²⁴ This is a narrower and more specific precedent than a general endorsement of “AI-generated evidence” in marketing-authorisation submissions; it qualifies one context-of-use for one histological scoring tool, and should not be read as establishing a blanket qualification pathway for AI protocol- or CRF-authoring tools.

The EU AI Act (Regulation 2024/1689) entered into force in August 2024; it establishes a risk-based framework in which “high-risk” status is not a blanket classification for all AI used in healthcare, but depends on specific criteria under Article 6: an AI system is high-risk either as a safety component of a regulated product (e.g., certain medical devices) or where it falls within an Annex III use case that poses a significant risk to health, safety, or fundamental rights.²⁵ Article 6(3) further provides a derogation for systems that perform a narrow procedural task, improve the result of a previously completed human activity, or require proper human review before their output is acted upon; criteria that may be relevant to the mandatory-expert-review design of the Document Generation and Metadata Automation tools discussed in this review, though the applicability of this derogation to any specific tool is a case-by-case legal determination outside the scope of this review and evolving via Commission guidance.²⁵

The revised ICH E6(R3) Good Clinical Practice guideline provides an additional normative anchor. Contrary to earlier drafts of this review, E6(R3) was not finalised in 2023: the ICH Assembly adopted the final guideline on 6 January 2025; it became applicable in the European Union on 23 July 2025; the US FDA issued its final E6(R3) GCP guidance in September 2025; and UK implementation occurred on 28 April 2026,illustrating that regional adoption timelines differ materially and should be checked against the applicable jurisdiction before citing a single “effective” date.²⁰ Its quality-by-design mandate—that every collected data element be traceable to a specific scientific or safety objective—creates a direct rationale for AI-assisted CRF rationalisation, since systematic traceability of this kind is difficult to achieve consistently through manual review alone (Table 3).