Development of consensus-based rules for AI-assisted Clinical Trial Agreement review: a Delphi study

Introduction

As the legal foundation of research operations, Clinical Trial Agreements (CTAs) must ensure compliance with regulatory mandates, ethical guidelines, and institutional policies. However, the review and negotiation of these agreements frequently represent critical bottlenecks, significantly delaying trial initiation (1). Such delays are further exacerbated by the increasing complexity of multi-center trials and the diverse legal frameworks across jurisdictions (2). Published estimates show that the average person-hours for a single CTA review (including negotiation and revision) is 80–120 hours for a single-center trial and 200–300 hours for a multi-center trial (3), with contract negotiation accounting for an average of 4–6 months of trial initiation delay (4). Although previous research has demonstrated that standardized clinical trial protocol templates can improve operational efficiency (5), traditional manual review processes remain heavily dependent on individual expertise, resulting in inconsistencies, elevated labor costs, and procedural inefficiencies, particularly in institutions managing high volumes of clinical research (3). Consequently, there is a pressing need for standardized, scalable tools to improve both the accuracy and efficiency of contract review workflows.

Prior studies and institutional audits (2,6) have quantified the most common CTA negotiation points, with the top five being: indemnification clauses (negotiation frequency: 92%), intellectual property ownership (88%), budget and payment terms (85%), publication rights (80%), and data ownership/storage (78%). These negotiation points are the primary sources of CTA review delays (accounting for about 70% of total negotiation time), which informs the domain focus of our rule set development.

The advent of artificial intelligence (AI) and large language models (LLMs) offers promising solutions for enhancing expertise and boosting productivity for contract management. AI-driven contract analysis tools have shown considerable potential to streamline review processes and improve risk management strategies (7). Nevertheless, despite recent technological advancements, general-purpose LLMs encounter notable limitations in the precision-driven context of contract review. These models often lack domain-specific knowledge, resulting in an inability to grasp nuanced legal meanings, omission of critical interdependencies between clauses, and ultimately compromising the reliability of contract interpretations (8). Existing AI methodologies frequently rely on superficial semantic or lexical similarity measures, which are insufficient for navigating the intricate, context-dependent nature of legal and technical documents. Moreover, the unstructured and voluminous format of contractual materials presents additional challenges for effective information retrieval, further diminishing system efficacy. Addressing these challenges necessitates a specialized approach that systematically integrates expert domain knowledge into the LLM or retrieval-augmented generation (RAG) framework.

Although previous studies have highlighted critical risk domains and essential contract language in CTAs, a fully integrated framework that draws on multidisciplinary expertise is still lacking. More importantly, there is no established methodology for converting the extensive administrative knowledge in contract management into structured, machine-interpretable formats that AI systems can effectively utilize. This study addresses this core gap by developing a consensus-based, LLM-ready rule set tailored for CTA review—representing a foundational step in bridging unstructured expert experience with AI operationalization.

The Delphi method is a structured communication and consensus-building approach and has been widely validated in healthcare research. It provides a robust mechanism for synthesizing expert insights (9). Its proven use in clinical trial management—such as creating quality assessment frameworks (10) and risk mitigation protocols (11)—demonstrates its capacity to balance scientific rigor with practical feasibility. Building on this foundation, our study applies the Delphi method to develop a consensus-driven, expert-validated rule set that is explicitly designed for subsequent AI integration.

This study aimed to establish a comprehensive set of LLM-ready rules to guide LLM -assisted CTA review process, covering legal, financial, and operational aspects, tailored specifically to China’s regulatory environment. We hypothesized that leveraging expert insights from diverse Chinese institutions would enable the creation of a rule set that is both legally aligned with Chinese regulatory requirements and operationally practical—serving as a critical precursor for future AI system development and validation.

Methods

Study design

This Delphi process was conducted in three rounds, following the guidelines outlined in the Conducting and Reporting of Delphi Studies (CREDES) and reporting quality indicators (12). The study was conducted from March 2025 to May 2025.

This study focuses on analyzing the legal validity of clinical trial contract clauses, optimizing contract review workflows, and comparing differences across industry-standard contract templates. Its research objects are contract texts, commercial provisions, and legal provisions themselves—not human subjects or sensitive personal data (e.g., health information, biological samples, personal behavioral data). No collection, use, or disclosure of individual-level sensitive information is involved; the Delphi survey only solicits experts’ professional judgments on regulatory rules and contract clauses, with all responses anonymized and unrelated to personal privacy.

Development of the preliminary rules

A preliminary set of contract review rules was created through a structured three-step approach. First, an extensive and systematic literature review was undertaken in PubMed, CNKI, WanFang, and the ICH official database for the period January 2020 to February 2025, with core keywords including “Clinical Trial Agreement review”, “ICH-GCP”, “Chinese GCP”, “contract review rules”, and “AI-assisted contract analysis”. The review was based on the major regulatory frameworks including the International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use-Good Clinical Practice (ICH-GCP) (13) and the Good Clinical Practice (GCP) (14) guidelines issued by China’s National Medical Products Administration (NMPA). Additionally, institutional Standard Operating Procedures (SOPs) sourced from 12 NMPA-compliant tertiary general and oncology specialty institutions with ≥5 years of CTA review experience were examined, and only formally issued and implemented SOPs were included to determine the key contractual components for legal and operational compliance. Second, inputs were provided by five clinical trial administrators with over 2 years of CTA review experience, who were recruited from the above-mentioned tertiary institutions across different cities. The five clinical trial administrators were from 5 different tertiary institutions—no single-institution bias and they represent diverse trial types and key therapeutic areas, ensuring their inputs cover the practical pain points of CTA review across different trial types and therapeutic areas.

Finally, the study team adopted thematic synthesis for consolidation: regulatory requirements and practical insights were coded into pre-defined themes, with overlapping rules merged and ambiguous content refined via team unanimous agreement (discrepancies resolved by an independent senior researcher with over 15 years of clinical trial contract and regulatory compliance experience). The consolidated preliminary rules were standardized with three core elements—clear review criterion, definite non-compliance consequence, and designated applicable party—and organized into formal and substantive review categories, covering legal/regulatory compliance, financial provisions, risk management, intellectual property, and quality assurance. Most rules align with established standards, while some are tailored to institutional needs and operational priorities.

Participants enrollment criteria

The whole participants represented a diverse group of professionals, including legal advisors and contract managers from sponsoring organizations, seasoned lawyers specializing in CTA law, and clinical research managers and investigators from tertiary general hospitals and oncology specialty institutions across China. To qualify, the participants should have a minimum of three years of experience in CTA review or management and had to provide voluntary consent. Individuals with potential conflicts of interest—such as consultants for AI-based contract review platforms—were excluded to ensure objectivity and credibility. Each participant should have extensive experience in contract review, ethical compliance, and GCP standards. To contextualize their contributions, demographic and professional details, such as roles, institutional affiliations, and years of relevant experience, were documented. Participants were recruited from different regions across China, including cities such as Guangzhou, Shanghai, Beijing, Kunming and Fuzhou, to ensure the geographic diversity. Consistent with the established recommendations for Delphi studies, which advise 15 to 50 participants (15), we aimed to recruit at least 20 participants each round to ensure heterogeneity and comprehensive domain insight.

Delphi process

The Delphi method collects expert feedback through questionnaires instead of face-to-face meetings, preserving anonymity and minimizing bias from dominant voices. In the first round, 23 participants received an email containing a three-part questionnaire: (I) an overview of the survey, (II) a table that listed the 74 initial rules alongside columns for the Likert scale, importance classification, and free-text comments to capture their specific recommendations or rationale for change, (III) a section for personal information and an optional section for additional suggestions, where participants could provide any extra rules or feedback at their discretion.

Participants assessed each rule using a five-point Likert scale from 1 (strongly disagree) to 5 (strongly agree) and classified its importance as “critical”, “important” or “moderate”. A free-text comment was mandatory to justify any rating of 3 or below. Additionally, participants rated their familiarity with the content on a five-level scale (5= familiar, 1= not familiar) and indicated the basis of their judgement, by selecting from 4 categories: theoretical analysis, practical experience, reference to relevant literatures or intuitive judgment. To maintain the quality and timeliness, participants were asked to complete the questionnaire within ten days, with reminders sent by email or telephone if necessary.

There is no universally accepted definition of consensus in Delphi studies (16), and approaches vary considerably across literature. Commonly used methods include percentage agreement, measures of central tendency, and dispersion metrics. To address the limitations of single metrics—such as the sensitivity of the mean to outliers or the potential for the mode to mask bimodal distributions—this study applied a comprehensive, pre-defined set of criteria for evaluating consensus. Rules were categorized as low consensus if they had a coefficient of variation (CV) >25%, a mean score <3.5, or a standard deviation (SD) >1.2. Conversely, rules with a mean score >4.5 and an SD <0.8 were classified as high consensus. Remaining rules were classified as moderate consensus.

The ≥85% threshold for “critical/important” ratings is based on recommended standards for Delphi studies in healthcare research (16,17) and is widely used in consensus development for clinical practice guidelines and regulatory rules. This threshold ensures that the rules have strong expert endorsement and avoid the inclusion of rules with significant expert disagreement, which is critical for the reliability of AI-assisted review rules that require high consistency. A post-hoc sensitivity analysis was conducted: (I) if the threshold was lowered to 75%, 8 additional rules would be included (all with medium consensus), but these rules had >15% expert ratings of “moderate” importance, indicating insufficient practical necessity; (II) if the threshold was raised to 95%, 6 rules with 85–94% “critical/important” ratings would be excluded, most of which were operational details (e.g., bilingual contract execution), leading to an over-restrictive rule set. Thus, the 85% threshold balances expert consensus and practical comprehensiveness.

Regarding CV threshold sensitivity: (I) CV >20% (stricter): 12 additional rules classified as low consensus (excluded); (II) CV >30% (looser): 8 fewer rules classified as low consensus (included). The chosen CV >25% threshold balances stringency and comprehensiveness.

Rules in the low-consensus category were retained, revised or removed based on the research team’s decisions, which were informed by feedback from the expert panel. Meanwhile, rules in the moderate or high-consensus categories, which fewer than 85% of participants rated as ‘critical’ or ‘important’, were also modified based on qualitative feedback. New rules suggested by over two participants would be adopted by the research team and incorporated into the updated set. Subsequent Delphi rounds re-evaluated these revised and newly-added rules using the same scoring criteria. After each round, participants received a summary of prior results to encourage reflection and support consensus development. This iterative process continued until all rules simultaneously meet the following two conditions: (I) high or medium consensus; (II) cumulative proportion of importance rated as “important” and “critical” ≥85%.

The Delphi survey was conducted using an online platform (www.wjx.cn) in mainland China, which enables real-time data collection and analysis, while maintaining anonymity and reducing potential bias.

Quality control

All returned questionnaires were carefully reviewed for completeness and accuracy. Incomplete responses were excluded from further analysis. Upon collection of all completed forms, the open-ended comments were independently reviewed by the study team. The study team then systematically reviewed and analyzed these qualitative inputs to guide the modification, consolidation, or elimination of items, ensuring the alignment with the study’s objectives and supporting literature.

Statistical analysis

Data were analyzed using SPSS (version 26.0; IBM Corp., Armonk, NY, USA) for analysis. Descriptive statistics included frequency, proportion, mean, SD and CV. Participant authority was quantified by the composite reliability (Cr), calculated as Cr = (Cs + Ca)/2, where Cs, represents the Familiarity Score and Ca the Judgment Basis Score. Cs was assessed via a 5-point Likert scale from 1 (“not familiar”) to 5 (“familiar”) and normalized to a 0–1 range. Ca reflected the primary basis for expert judgment, weighted as follows: theoretical analysis (0.3), practical experience (0.5), peer recommendations (0.1), and intuitive judgment (0.1). Both Ca and Cs were self-reported by the participants. A Cr value ≥0.70 indicated the high authority (12). Participant engagement was expressed through questionnaire recovery rates, while consensus consistency was measured using Kendall’s W, ranging from 0 to 1, with higher values signifying stronger agreement.

Ethical consideration

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. No IRB approval and consent forms are needed as this study didn’t involving human biomedical research or sensitive personal information processing.

Results

Demographics of participants

Twenty-three participants were enrolled in the Delphi Process. Table 1 presents the expert demographic data. The sample comprised one participant with Ph.D. degree, 13 with master’s, and 9 with bachelor’s degrees. In terms of composition, professional experience, and areas of specialization, the participants met the principles of diversity and representativeness, providing constructive feedback and valuable suggestions for the study.

Positivity, authority, and coordination of experts

The Delphi survey achieved reconciliation rates of 100% in the first two rounds (23/23) and 95.65% in the third (22/23), reflecting strong engagement and interest among experts. The mean Cr across all rounds was 0.89, indicating a high level of authority on the subject matter. After three rounds, all items demonstrated a CV ≤0.25 and Kendall’s W ranging from 0.87 to 0.91 (P<0.001), confirming a high degree of consensus among the participants.

Delphi consensus outcomes

Figure 1 illustrated the Delphi survey outcome. In Round 1, all 23 participants evaluated the initial set of 74 contract review rules, with 78.3% (18/23) suggesting revisions, generating 155 actionable comments. Feedback primarily addressed linguistic precision (52.9%), structural optimization (28.4%), and content expansion (18.7%). Based on predefined consensus criteria, 14 initial rules required reassessment—nine with medium consensus and five with low consensus—while 60 rules achieved immediate high consensus. The study team incorporated some participants’ recommendations without removing any rules. As shown in Table S1, the study team revised and refined specific rules, which included: adding the description “if the required number of participants is reached and additional participants need to be enrolled, a supplementary agreement to increase the number of participants must be signed” to Rule X16; revising the wording in Rule X46 from “if the other party discovers any new uses of investigational products eligible for patenting shall be co-owned” to “if the other party discovers any technical features or design that meet the criteria for patent application, the patent application rights and patent rights shall be negotiated separately by both parties” in order to standardize terminology and align with patent law provisions; modifying the time frame in Rule X48 from “ten working days” to “within the agreed time period” to provide greater flexibility; deleting the specific requirement “the monitoring frequency should be ≥1 time/month” from Rule X67; adding more detailed scenario descriptions to Rule X72; and introducing two new rules X75 (about liability: “Explicit penalty clauses for sponsor protocol deviations”) and X76 (about administrative: “Bilingual (English/Chinese) contract execution”).

Figure 1 Flow diagram of the Delphi survey process.

In Round 2, the updated set of 16 rules (14 initial +2 new rules) were re-assessed by the original 23 participants, through a blinded review to minimize bias. Participants also received the scoring statistics for the 14 initial rules from the first round of the survey, along with a comparative indication of any modifications made to the rules or notes stating that no modifications were made, to be used as reference during the second round of evaluation. Overall, 60.87% (14/23) of participants submitted 36 substantive comments. Of these, 86.1% (31 comments) supported the updated set of 16 rules, while 13.9% (5 comments) raised concerns about two specific rules, X4 and X15. Based on predefined consensus criteria, these two rules required reassessment. Rule X4, addressing tax refund procedures, was initially stated as “Sponsor shall bear non-refundable taxes during trial fund reimbursement”. Participants noted this conflicted with national tax policies, particularly value-added tax (VAT) refund eligibility under China’s Tax Collection Law. The rule was revised to “Upon sponsor’s application to tax authorities, the institution may refund taxes after receiving official approval”. For Rule X15, which involving protocol numbers in contracts, the study team believed that it is necessary to retain the original wording, as including protocol numbers could risk confidentiality breaches according to ICH GCP E6(R3) (13).

In Round 3, all 22 participants evaluated the two disputed rules, and consensus was achieved for both. Rule X4 received a mean score of 4.2 (SD =0.6, CV =14.3%), while Rule X15 scored 4.5 (SD =0.5, CV =11.1%).

The list of final rules, as shown in Table 2 and Table S1, consists of 76 validated items after three rounds of Delphi, representing a comprehensive, standardized approach to contract review. Table 2 presents the core final rule set organized by rule category, applicable party, and specific final version rules. Table S1 provides the full revision history. The statistical results of expert evaluations of the rules during the three rounds of Delphi process are shown in Table S2. The inclusion of expert-driven revisions ensures the rules are practical, legally aligned with Chinese regulatory requirements, and adaptable to evolving clinical trial environments.

Discussion

This Delphi study tackles a pivotal challenge in clinical trial management: the absence of standardized, scalable approaches for contract review. To address this, it systematically converts expert administrative knowledge into a structured rule set tailored for LLM-assisted implementation. Unlike general-purpose AI contract review tools (7), which often operate as non-transparent “black boxes” and fail to capture domain-specific legal nuances, our consensus-driven framework delivers transparent, regulatory-aligned rules that serve as a reliable knowledge base for LLM systems.

A key distinction from existing work is as follows: most existing CTA templates/checklists (4) are unstructured or semi-structured and designed for human review; our work is the first to develop a structured, LLM-ready rule set with standardized terminology and logical structures that LLM can parse and execute. Additionally, our rule set is consensus-based (23 experts, 3-round Delphi) and tailored to China’s regulatory context. Compared to prior AI-assisted contract review systems (7,8), which use end-to-end LLM reasoning and lack domain-specific CTA knowledge, our system combines a consensus-based rule set with LLM reasoning, improving review accuracy and consistency for CTAs. Unlike regulatory automation frameworks that focus on single domains (e.g., data management), our rule set covers all core CTA review domains (legal, financial, operational, quality management).

One of the most significant findings was the strong emphasis of participants on linguistic precision in contract drafting. More than half of the feedback (52.9%) focused on improving terminological clarity and ensuring alignment with relevant legal frameworks, such as patent law. For example, revising the rule on “new uses of investigational products eligible for patenting” to conform with patent law definitions highlights the importance of legal and technical accuracy in contract language. This attention to precision is critical for minimizing ambiguities and the enforceability of CTAs.

The inclusion of penalty clauses for sponsor protocol deviations and bilingual contract execution responds to two key challenges in clinical trial management: the growing complexity of international trials and the need for clear, accessible legal language for all stakeholders (5). The inclusion of these rules demonstrates the ability of the Delphi method to identify and address practical issues that may not be fully captured in existing contract review guidelines.

The study highlights the effectiveness of the Delphi method in establishing consensus on rules that harmonize scientific rigor with practical feasibility. Such balance ensures that the final rule set is not only legally aligned with Chinese regulatory requirements but also adaptable to the complex, evolving landscape of clinical trials. The high level of consensus and strong endorsement of rule importance indicate that the resulting framework aligns well with expert expectations and provides a reliable foundation for LLM-assisted contract review agents.

A major conceptual consideration is the role of fixed rule sets in the era of LLMs and generative AI. The rule set is not a rigid replacement for LLM contextual reasoning but a high-fidelity knowledge base to constrain and guide LLM reasoning. LLMs have strong contextual understanding but suffer from hallucinations and inconsistent judgment in domain-specific tasks (e.g., CTA review); the rule set provides standardized, regulatory-aligned constraints to reduce hallucinations and ensure AI review results are consistent with expert and regulatory requirements.

The rule set offers several advantages over end-to-end LLM reasoning, fine-tuning, or reinforcement learning from human feedback (RLHF): (I) transparency: rule-based review is interpretable (AI can clearly flag which rule a CTA violates), whereas end-to-end LLM reasoning is a “black box”; (II) regulatory compliance: clinical trial regulation requires traceable review decisions, which rule-based sets provide; (III) low cost of update: The rule set can be modularly updated for new regulatory requirements, whereas LLM fine-tuning/RLHF requires large amounts of labeled data and computational resources.

Rules may become obsolete, brittle, or counterproductive under the following conditions: (I) rapid regulatory change without corresponding rule updates; (II) emerging trial designs with no corresponding rule adaptations (e.g., AI-driven adaptive trials); (III) over-reliance on rules without LLM contextual reasoning (e.g., rigid application to non-standard CTAs with unique clauses). To avoid this, we propose a human-in-the-loop + modular rule update mechanism to combine rule-based constraints with LLM flexibility.

The expert consensus achieved in this Delphi study provides a robust set of natural language rules, raising an important question: how can these rules be operationalized for an AI-assisted review process? While the rules are not AI-executable, they are LLM-utilizable-designed to be AI-ready, serving as a validated, high-fidelity knowledge base for building reliable agents. Simple checks—such as verifying the presence of key clauses—can be handled through pattern-matching techniques like regular expressions. More complex rules, involving logical inference (e.g., cross-verifying protocol and budget consistency or assessing regulatory compliance), would be encoded within a rule engine or integrated into structured prompts for LLMs using a RAG framework. This enables a human-in-the-loop AI agent that flags potential violations, highlights associated risks, and provides expert rationale without autonomously approving or rejecting contracts. Ultimately, this study lays the foundation for developing transparent, efficient AI-assisted review agent that improve consistency and speed while preserving essential human oversight.

Although this study established a foundational set of LLM-utilizable rules for CTA review, we acknowledge that formal, large-scale AI validation and comparative testing have not been fully completed at this stage. However, we have preliminarily input the 76 consensus rules into a LLM and conducted initial AI-assisted review on real CTA documents. We plan to further verify each rule manually, compare LLM performance with expert review in terms of issue identification efficiency and time consumption, and will report these validation results in detail in subsequent studies.

Regarding the applicability to diverse clinical trial designs: the rule set is primarily design-agnostic and based on core regulatory requirements (ICH-GCP, Chinese GCP) that apply to all clinical trial types. The core domains (legal compliance, financial provisions, IP ownership, data management, safety reporting) are universal for all trial designs, and the rules in these domains are fully applicable to adaptive designs, decentralized/hybrid trials, and platform studies.

This study has several limitations. The expert panel was composed primarily of professionals from China [73.91% clinical research management, 13.04% contract management of sponsor or Contract Research Organization (CRO), 4.35% legal compliance, 8.70% medical ethics], which may limit the generalizability of the developed rules to other regulatory contexts. While the proportional representation of specialized legal and ethics experts is relatively low, it is important to note that all 23 participating experts possess ≥3 years of practical CTA review experience and have received systematic training or hold educational backgrounds in clinical trial-related legal compliance. Specifically, all experts completed mandatory training on core legal frameworks including China’s Patent Law, Medical Administration Law, and ICH-GCP/Chinese GCP regulatory requirements—key foundations for evaluating legal clauses in CTAs. Their day-to-day CTA review work involves frequent assessment of legal-critical provisions (e.g., indemnification, intellectual property, tax compliance), equipping them with practical legal insight aligned with the operational context of CTAs. To mitigate potential gaps in legal domain depth (e.g., patent law alignment for intellectual property clauses, tax policy compliance for financial clauses), we supplemented the expert consensus with targeted measures: (I) all legal-related rules were cross-verified against the latest Chinese laws and regulations to ensure regulatory alignment; (II) two external legal experts specializing in pharmaceutical clinical trials were consulted to validate the legal precision of core clauses. Future research will expand the expert panel to include more legal, ethics, and CRO professionals (targeting 20–30% of the panel) to further enhance the depth of legal domain coverage and enrich multi-perspective insights for rule refinement.

Conclusions

This study represents the first effort to combine Delphi-derived expert consensus with AI operationalization for CTAs in this context. The resulting framework aligns with international standards and local regulatory requirements, offering a relevant and scalable solution for CTA management in China. The Delphi methodology offers a transferable blueprint for developing similar LLM-assisted agents in other jurisdictions.

Acknowledgments

We would like to express our gratitude to our expert colleagues who provided professional opinions in the Delphi survey.

Footnote

Data Sharing Statement: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-1-255/dss

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-1-255/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-1-255/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. No IRB approval and consent forms are needed as this study didn’t involving human biomedical research or sensitive personal information processing.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Dilts DM, Sandler AB. Invisible barriers to clinical trials: the impact of structural, infrastructural, and procedural barriers to opening oncology clinical trials. J Clin Oncol 2006;24:4545-52. [Crossref] [PubMed]
2. Thompson MA, Hurley PA, Faller B, et al. Challenges With Research Contract Negotiations in Community-Based Cancer Research. J Oncol Pract 2016;12:e626-32. [Crossref] [PubMed]
3. Pfeiffer JP, Richmond FJ. Management of Clinical Trial Agreements: Current Practices of Investigators in the United States. Ther Innov Regul Sci 2015;49:139-45. [Crossref] [PubMed]
4. Lai J, Forney L, Brinton DL, et al. Drivers of Start-Up Delays in Global Randomized Clinical Trials. Ther Innov Regul Sci 2021;55:212-27. [Crossref] [PubMed]
5. Lawrence CE, Bruce VNM, Salberg LD, et al. Quantitative assessment of the impact of standard agreement templates on multisite clinical trial start up time. J Clin Transl Sci 2023;7:e204. [Crossref] [PubMed]
6. Pfeiffer JP. Clinical trial contracts: Language and limitations. 2017;18. Available online: https://www.pharmoutsourcing.com/Featured-Articles/336113-Clinical-Trial-Contracts-Language-and-Limitations/
7. Giampieri P. AI-Powered Contracts: a Critical Analysis. Int J Semiot Law 2025;38:403-20. Available online: https://doi.org/10.1007/s11196-024-10137-z
8. Kim EW, Shin YJ, Kim KJ, et al. Development of an Automated Construction Contract Review Framework Using Large Language Model and Domain Knowledge. Buildings 2025;15:923. Available online: https://doi.org/10.3390/buildings15060923
9. Boulkedid R, Abdoul H, Loustau M, et al. Using and reporting the Delphi method for selecting healthcare quality indicators: a systematic review. PLoS One 2011;6:e20476. [Crossref] [PubMed]
10. Elwyn G, O’Connor A, Stacey D, et al. Developing a quality criteria framework for patient decision aids: online international Delphi consensus process. BMJ 2006;333:417. [Crossref] [PubMed]
11. Hafeez S, Shakil-Ur-Rehman S, Riaz S, et al. Consensus-Driven Development of an Exercise Base Manual Programme for Prediabetic Patients: A Delphi Study. J Multidiscip Healthc 2025;18:2461-76. [Crossref] [PubMed]
12. Dai F, Wei K, Chen Y, et al. Construction of an index system for qualitative evaluation of undergraduate nursing students innovative ability: A Delphi study. J Clin Nurs 2019;28:4379-88. [Crossref] [PubMed]
13. International Council for Harmonisation. ICH E6(R3): Guideline for Good Clinical Practice. 2025. [Last accessed on 2025 Feb 09]. Available online: https://database.ich.org/sites/default/files/ICH_E6%28R3%29_Step4_FinalGuideline_2025_0106.pdf
14. Yang F, Heng J, Li K, et al. The new Good Clinical Practice-2020 in China: Views from ethical perspective. Lancet Reg Health West Pac 2021;8:100117. [Crossref] [PubMed]
15. McPherson S, Reese C, Wendler MC. Methodology Update: Delphi Studies. Nurs Res 2018;67:404-10. [Crossref] [PubMed]
16. Shang Z. Use of Delphi in health sciences research: A narrative review. Medicine (Baltimore) 2023;102:e32829. [Crossref] [PubMed]
17. Nasa P, Jain R, Juneja D. Delphi methodology in healthcare research: How to decide its appropriateness. World J Methodol 2021;11:116-29. [Crossref] [PubMed]