How to Qualify AI Functions in Clinical Trial Software
TL;DR
FDA's drug-development AI credibility guidance remains a draft [1]. This guide applies its context-of-use questions to individual trial software functions and proposes four controls.
- Identify binding regulations and applicable guidance for the specific use [1] [2] [3] [4] [5] [6] [7] [8].
- Rate the function and confirm who makes the decision. Reject unconfirmed automated writes to regulated trial records unless the function is expressly qualified and authorized under the highest-control tier (Tier 4) for that context of use.
- Specify the approved model and inputs. Test changes using a fixed set with known answers. Follow applicable EMA expectations for pivotal analysis [9].
- Preserve the proposed output and the human decision in the audit evidence. Use review signals to select cases for inspection.
One deployment study recorded 18 institutions activating REDCap AI. At Vanderbilt, 958 projects generated calls for writing, summarization and translation [10]. Those workflows give examples of use under administrator controls. Evaluate clinical functions separately.
1. What Binds, What Guides, and What Is Still Draft (as of Late 2026)
Start by identifying the authority, version and intended use relevant to the function. Chart 01 separates binding rules, adopted guidance and drafts. EClinCloud's proposed controls start in Section 3.
1.1 United States: Five Regulatory and Guidance Baselines
Sponsors conducting clinical trials under Investigational New Drug (IND) or Investigational Device Exemption (IDE) regulations interact with five distinct FDA tracks governing software systems, artificial intelligence, and electronic records:
- FDA drug and biological product AI draft. FDA issued the January 2025 draft for comment. Section II covers AI used to produce information or data intended to support regulatory decisions about drug safety, effectiveness or quality. It leaves out operational efficiencies that do not impact patient safety, drug quality or the reliability of results from a nonclinical or clinical study. Its seven steps are: define the question, define context of use, assess model risk, plan credibility evidence, execute the plan, document results and deviations, and determine adequacy for the context. Model influence and decision consequence are the two factors in step 3. Use formal meetings such as INTERACT or Pre-IND, or a Table 1 option such as C3TI where appropriate [1].
- Transatlantic Good AI Practice Guiding Principles (Non-Binding Joint Principles): In January 2026, FDA and the European Medicines Agency (EMA) published 10 joint guiding principles for AI in drug development [11]. The official headings are: (1) Human-centric by design; (2) Risk-based approach; (3) Adherence to standards; (4) Clear context of use; (5) Multidisciplinary expertise; (6) Data governance and documentation; (7) Model design and development practices; (8) Risk-based performance assessment; (9) Life cycle management; and (10) Clear, essential information [11]. These principles represent non-binding consensus recommendations intended to inform sponsor adoption and regulatory thinking rather than independent legal mandates.
- Final Electronic Systems, DHT, and CSA Guidances (Non-Binding Recommendations Interpreting Binding 21 CFR Rules): Agency guidances explicitly state that they represent FDA's current thinking and offer non-binding recommendations unless specific statutory or regulatory provisions are cited. Under FDA's October 2024 final guidance, Electronic Systems, Electronic Records, and Electronic Signatures in Clinical Investigations: Questions and Answers (Revision 1) [12], the agency interprets binding requirements under 21 CFR Part 11 (Electronic Records; Electronic Signatures) and 21 CFR Part 312 (Investigational New Drug Application). Key recommendations include generating verified certified copies that maintain the same information, context, content, and structure as original records (Q3), retaining records in human-readable formats with essential metadata (Q5), implementing risk-based validation for intended purposes (Q7), maintaining secure, computer-generated, time-stamped audit trails that capture user actions without overwriting prior entries (Q12), capturing deliberate user actions such as saving or submitting data rather than background keystrokes (Q13), and establishing secure system clock controls that document the date, time, and time zone or Greenwich Mean Time (GMT) (Q14) [12]. In parallel, FDA's December 2023 final guidance, Digital Health Technologies for Remote Data Acquisition in Clinical Investigations, recommends identifying the minimum technical and performance specifications of a DHT and the specific model or version where applicable (section IV.A.2). It also recommends recording when and how each DHT was updated and confirming that it remains fit for purpose after the update (section IV.H.4) [13]. Furthermore, on February 3, 2026, FDA reissued its final guidance, Computer Software Assurance for Production and Quality Management System Software (Docket FDA-2022-D-0795), aligning validation terminology with the Quality Management System Regulation (QMSR, 21 CFR Part 820, effective February 2, 2026) [14]. Although Part 820 directly binds medical device manufacturers rather than pharmaceutical trial sponsors as such, the CSA guidance's distinction between scripted testing and unscripted testing (including exploratory testing) provides a flexible, risk-based testing framework that sponsors can adapt for eClinical software qualification [14].
- Investigational Device Software Rules vs. Post-Market Change Control Plans: When software functions meet the definition of a device under Section 201(h) of the Federal Food, Drug, and Cosmetic Act (FD&C Act), investigational use in clinical trials is governed by 21 CFR Part 812 (Investigational Device Exemptions) [13]. For investigations conducted under an approved IDE application, protocol and software changes implemented during the active study must be evaluated under 21 CFR 812.35 to determine whether a supplemental IDE submission is required before implementation, whereas studies conducted under abbreviated requirements (812.2(b)) or exempt investigations (812.2(c)) adhere to their respective regulatory pathways. By contrast, FDA's August 18, 2025 reissued final guidance, Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions (Docket FDA-2022-D-2628), represents an optional marketing-submission pathway for commercially authorized devices evaluated under 510(k), De Novo, or Premarket Approval (PMA) pathways [15]. A Predetermined Change Control Plan (PCCP) allows a manufacturer to pre-specify post-market model modifications across three components: a Description of Modifications, a Modification Protocol, and an Impact Assessment [15]. A PCCP is not an authorization mechanism for mid-study algorithm modifications during an investigational drug or device trial, and its companion document, Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations (Docket FDA-2024-D-4488, issued January 7, 2025), remains Draft Guidance — Not for Implementation [16].
- FDA Final Guidance on ICH E6(R3) (Level 1 Non-Binding Guidance): In September 2025, FDA published E6(R3) Good Clinical Practice (GCP) as final Level 1 guidance under Docket FDA-2023-D-1955 [4]. In the United States, ICH guidelines adopted by FDA represent agency guidance containing non-binding recommendations rather than direct federal statutes. Sponsors submitting trial data to CDER or CBER under an IND can reference E6(R3) concepts as reflecting agency expectations, and binding obligations come from the applicable regulations. The September 2025 guidance predates the final ICH Annex 2. FDA’s Annex 2 page still labeled its version a draft when checked on November 13, 2026 [17].
1.2 ICH E6(R3), European Union, United Kingdom, and China Frameworks
International regulatory frameworks distinguish between directly binding regional regulations, adopted consensus guidelines, agency reflection papers, and national decrees:
- ICH E6(R3) Consolidated Guideline (Step 4 Adopted June 16, 2026): The International Council for Harmonisation (ICH) consolidated E6(R3) at Step 4 on June 16, 2026, uniting the overarching Principles, Annex 1 (interventional clinical trials, finalized January 6, 2025), and Annex 2 (additional considerations for non-traditional trial designs, decentralized trials, and real-world data sources, adopted June 3, 2026) [2]. Under ICH operating procedures, Step 4 guidelines are recommended for adoption to regional regulatory authorities rather than serving as directly binding international law. In the European Union, the EMA Committee for Medicinal Products for Human Use (CHMP) adopted the Principles and Annex 1 with an effective implementation date of July 23, 2025, while Annex 2 and the consolidated text carry an official effective date of January 15, 2027 [3]. In jurisdictions where implemented, Section 4.2.2 states that computerised systems should maintain audit trails that capture system-user interactions and workflow modifications with an unambiguous date and time (giving Coordinated Universal Time, UTC, as an example); Section 4.2.3 states that sponsors should conduct risk-based reviews of data and metadata, including audit trails; and Section 4.3.4 states that computerized system validation should be proportionate to risk, covering system configuration, integration, and changes over time [2].
- European Union Clinical Trials Regulation and EMA Guidelines: In the EU, clinical trials on medicinal products are governed by Regulation (EU) No 536/2014 (Clinical Trials Regulation, CTR) [18]. The CTR became applicable on January 31, 2022. Article 47 requires the sponsor and investigator to take appropriate account of ICH good clinical practice guidelines when drawing up the protocol and applying the Regulation and the protocol [18]. Article 98 provided transitional treatment for trials under the previous Directive. When inspecting a historical trial, determine the applicable regime from its authorization and transition history [18]. Operational expectations for computerized systems are set forth in EMA's GCP Inspectors Working Group guideline, Guideline on computerised systems and electronic data in clinical trials (EMA/INS/GCP/112288/2023, effective six months after publication in March 2023) [19]. Its scope includes AI used for participant recruitment, eligibility, coding, data clarification, query processes and event adjudication. The general expectations apply to these uses. The guideline reserves AI-specific requirements for a possible future annex [19]. This guideline recommends formal vendor qualification, clear role segregation between administrative personnel and trial users, and complete chronological exportability of audit logs [19]. In September 2024, EMA adopted its non-binding Reflection paper on the use of artificial intelligence (AI) in the medicinal product lifecycle (EMA/CHMP/CVMP/83833/2023) [9]. Section 2.3.3.1 says that where a method has high regulatory impact or high patient risk and EMA has yet to qualify it for that context, EMA may request the model architecture, development and testing records, training data and the processing pipeline. Section 2.3.3.3 addresses AI used in trial data analysis and inference, including controls for late-stage pivotal trials [9]. In parallel, the European Commission's public consultation on revised EudraLex Volume 4 Chapter 4, Annex 11 (Computerised Systems), and draft Annex 22 (Artificial Intelligence) closed in October 2025 [20]. When checked on November 13, 2026, the Commission’s EudraLex Volume 4 still listed the 2011 Chapter 4 and Annex 11 and contained no Annex 22. The revised texts remained GMP drafts. Clinical trial operations are governed separately by the applicable GCP requirements.
- United Kingdom MHRA AI Airlock Sandbox: The UK MHRA, supported by the Department of Health and Social Care (DHSC) and the Regulatory Innovation Office (RIO), published its AI Airlock Sandbox Phase 2 Programme Report in June 2026 (updated July 2026) [21]. The Airlock is a regulatory sandbox for AI medical devices. Its report sets no GCP requirement for trial software. The Phase 2 sandbox evaluated 51 applications, selecting seven pilot candidate projects (three tested across multi-environment pathways and four evaluated in simulation-only workshops) [21]. The report defined three distinct evaluation environments: Simulation environments (multidisciplinary stakeholder workshops stress-testing hypothetical risk scenarios), Virtual or Research environments (testing algorithms within secure, isolated retrospective data environments), and Real-World environments (deployment separate from clinical pathways and decision-making processes, utilizing retrospective or synthetic data, which does not constitute a clinical investigation) [21]. The report highlighted two key lifecycle insights: Insight 2 observed that human oversight can degrade over a system's lifecycle as clinicians develop automation complacency, while Insight 3 emphasized that for large language model (LLM) products, significant shifts in behavior can occur through model behavior alone, such that systems lacking active guardrails may exhibit functions beyond their intended purpose [21]. MHRA launched Phase 3 in October 2026 (running through April 2029) to focus specifically on post-market surveillance and lifecycle regulatory oversight [22].
- China NMPA Revised Drug GCP and CDE Technical Guideline: Multi-regional trials operating in China must adhere to directly binding national regulation under NMPA Announcement 2026 No. 50, Good Clinical Practice for Drugs (2026 Revision), issued jointly by NMPA, the National Health Commission (NHC), the National Administration of Traditional Chinese Medicine (NATCM), and the National Disease Control and Prevention Administration (NDCPA) on May 21, 2026, and effective September 1, 2026 [5]. Announcement No. 50 repeals Announcement No. 57 of 2020 [5]. China’s CDE issued the Guideline on Computerized Systems and Electronic Data in Drug Clinical Trials under Notice 2026 No. 44 on September 22, 2026, effective on issuance [6]. The guideline describes the regulator’s current view and states that it has no mandatory legal force. It covers source-data location and direct capture, role-based access with privileged users independent of trial conduct, computer-generated and time-stamped audit trails with a minimum field list, supplier contracts, and qualification of cloud suppliers, including acceptance of audit or inspection. The guideline also says when a data point may be set to auto-fill [6]. See Section 5.1 of this guide.
- ISPE GAMP 5 Second Edition and GAMP AI Guide: In the engineering and validation domain, the International Society for Pharmaceutical Engineering (ISPE) published the GAMP Guide: Artificial Intelligence in July 2025 as a stand-alone guide intended for use alongside GAMP 5: A Risk-Based Approach to Compliant GxP Computerized Systems (Second Edition) [23]. The published table of contents covers concept, project, operation and retirement phases, quality risk management, supplier activities and management, critical thinking, data splitting, operational change and configuration management, and model monitoring and drift detection [23].
2026 Global Regulatory Status Matrix for AI/ML Across Clinical Trial Systems
Binding regulations, adopted guidelines and drafts for the United States, EU and China, with UK AI medical-device sandbox findings.
Scroll sideways for the full figure.
| Jurisdiction / Body | Regulatory Instrument & Formal Status | Effective / Application Date | Operational Trial Software Qualification Impact |
|---|---|---|---|
| United States (FDA) | AI in Drug Decision-Making (Docket FDA-2024-D-4689) — Draft Guidance (Not for Implementation) | Issued Jan 6, 2025; comments closed Apr 7, 2025 | FDA draft scope is defined in Section II. Formal meetings include INTERACT or Pre-IND. Table 1 includes other engagement routes such as C3TI. |
| United States (FDA) | ICH E6(R3) Good Clinical Practice (Docket FDA-2023-D-1955) — Final Level 1 Guidance | Published September 2025 | Final September 2025 guidance predates final ICH Annex 2. FDA still listed its separate Annex 2 as draft on November 13, 2026. |
| United States (FDA) | Electronic Systems & Signatures Q&A Rev 1 — Final Guidance (Non-Binding Recommendations) | October 2024 | Interprets binding 21 CFR Part 11; certified copy definitions (Q3), audit trail capture (Q12), and time zone or GMT clock controls (Q14). |
| United States (FDA) | Computer Software Assurance (CSA, Docket FDA-2022-D-0795) — Final Guidance (Non-Binding) | Reissued February 3, 2026 | Aligns with QMSR (21 CFR Part 820); distinguishes scripted testing from unscripted testing (including exploratory testing); adaptable for eClinical tools. |
| United States (FDA) | PCCP for AI Device Software (Docket FDA-2022-D-2628) — Final Guidance (Non-Binding) | Reissued August 18, 2025 | Optional post-market marketing submission pathway for commercial devices; does not authorize mid-trial algorithm changes in investigational IND/IDE studies. |
| United States & EU (FDA / EMA) | Guiding Principles of Good AI Practice in Drug Development — 10 Joint Principles (Non-Binding) | January 2026 | Ten overarching principles addressing human-centric design, standards adherence, lifecycle management, risk assessment, and data governance. |
| ICH | ICH E6(R3) Guideline for Good Clinical Practice — Step 4 Final Consolidated Guideline | Consolidated Step 4 adopted June 16, 2026 | ICH recommends unambiguous date and time, with UTC as an example, metadata review and risk-proportionate validation. |
| European Union (EMA) | ICH E6(R3) Regional Implementation — EMA CHMP Scientific Guideline | Principles & Annex 1: July 23, 2025; Annex 2 & Consolidated: January 15, 2027 | Adopted regional guideline in EU; establishes regulatory expectation for GCP computerized systems and risk-based audit trail governance. |
| European Union | Clinical Trials Regulation (EU) No 536/2014 — Directly Binding European Union Regulation | Applicable January 31, 2022, with Article 98 transition provisions | Applies to trials authorized under the Regulation. For older trials, determine the regime from the authorization and transition history. Article 47 requires taking appropriate account of ICH GCP guidelines. |
| European Union (EMA) | Computerised Systems in Clinical Trials (EMA/INS/GCP/112288/2023) — GCP IWG Guideline | Effective September 9, 2023 | General expectations cover AI for recruitment, eligibility, coding, data clarification, queries and event adjudication. EMA reserved AI-specific requirements for a possible future annex. |
| European Union (EMA) | AI Reflection Paper (EMA/CHMP/CVMP/83833/2023) — Non-Binding Reflection Paper | Adopted September 2024 | Section 2.3.3.3: EMA expects models and preprocessing used for hypothesis testing to be frozen and documented before database lock. |
| European Union | EU AI Act (2024/1689) & Digital Omnibus (2026/1744) — Directly Binding EU Regulations | AI Act in force August 1, 2024. Omnibus in force July 27, 2026. Article 50(1) applies August 2, 2026. | Article 50(1) disclosure where applicable. Chapter III Sections 1–3 except Article 6(5) apply to high-risk Annex III systems from December 2, 2027 and high-risk Annex I systems from August 2, 2028. |
| United Kingdom (MHRA) | AI Airlock — regulatory sandbox for AI medical devices | Phase 2 Report June/July 2026; Phase 3 launched Oct 2026 | Phase 2 selected 7 candidates from 51 applications. The report discusses automation complacency and sets no GCP requirement for trial software. |
| China (NMPA / NHC / NATCM / NDCPA) | Good Clinical Practice for Drugs (2026 Revision, Announcement 2026 No. 50) — Announcement (binding national GCP) | Issued May 21, 2026; Effective September 1, 2026 | Publishes the revised Good Clinical Practice for Drugs and repeals the 2020 announcement. |
| China (CDE) | Guideline on Computerized Systems and Electronic Data in Drug Clinical Trials (CDE Notice 2026 No. 44). The guideline states that it has no mandatory legal force. | Issued and effective September 22, 2026 | Direct source capture, role segregation and independence of privileged users, time-stamped audit trails, supplier contracts and qualifications, and limits on auto-fill. |
Binding regulations define legal obligations. The EU Clinical Trials Regulation also requires appropriate account of ICH GCP guidelines. The UK AI Airlock report describes a sandbox.
2. Published Evaluations, 2025–2026
The studies include a REDCap rollout, a retrospective support-ticket experiment, a site-risk model with external validation, a linguistic-validation study combining a review, survey and interviews, and two reviews of clinical use and regulatory records.
Shyr and colleagues studied the January–August 2025 rollout of REDCap 15.0 AI. Administrators at 18 institutions enabled the default-off features. Across 958 Vanderbilt projects, the features generated 2,801 writing, 2,720 summarization and 200 translation calls. The paper reports adoption and call counts. Output accuracy and effectiveness were outside its evaluation [10].
Carroll and colleagues evaluated large language models in managing user support inquiries at the University of Michigan [24]. Analyzing a sample of 90 support tickets drawn from 6,316 historical requests submitted during 2023, the investigators compared four GPT-4o configurations: default GPT-4o, GPT-4o supplemented with Vanderbilt documentation, GPT-4o with University of Michigan documentation, and GPT-4o with combined documentation [24]. The model demonstrated strong agreement with human expert categorization, achieving Cohen's kappa values of approximately 0.85 and categorization accuracy between 81% and 89% [24]. However, in the ticket complexity rating task, retrieval configurations over-rated ticket complexity; for draft response generation, the University of Michigan documentation configuration produced the most correct replies, generating substantially or fully correct answers on approximately two-thirds of routine low-to-moderate complexity tickets [24]. The authors concluded that LLMs show genuine promise for automated support triage when supported by ongoing human oversight, prompt refinement, and expanded local documentation [24].
Yang developed a site-qualification model with 460 sites from 42 ATTR-CM studies and used 761 sites from 89 DMD studies for external validation. A support vector machine performed best in the development setting. The study checked subgroups by geographic region and site type, and reported 81.21% accuracy in the external evaluation. Its DMD features used proxies. The study lacked ground-truth monitoring outcomes and left the projected operational effects untested in deployment [25]. EClinCloud recommends testing a proposed ranking against actual monitoring results before using it to allocate effort.
McKown and colleagues combined a literature review, interviews with AI specialists and a survey of 50 stakeholders. Respondents included instrument copyright holders, pharma COA or outcomes researchers, translation, eCOA and AI professionals, and authors of the earlier ISPOR work. Their recommendations weigh technical capability and efficiency alongside intellectual-property protection, data privacy and security, and patient involvement. They support AI with human oversight for most linguistic-validation steps and recommend excluding AI from reconciliation, clinician review and the conduct of cognitive debriefing interviews. They recommend human review at each step [26]. Karako and Gao's review reports evidence of equivalence to expert readers in automated left ventricular ejection fraction measurement. They identify generalization across vendors, subgroup fairness and algorithm changes during a trial as open challenges [27].
Thierry and colleagues reviewed 60 digital-health-technology-related approvals from September 2021 to March 2025. Electronic patient-reported outcomes tools accounted for 70% of the digital health tools studied. The authors found zero approvals based solely on a digital-health-technology-derived primary endpoint [28].
3. Classifying Each Function by Context of Use
Evaluate the function in its actual workflow. A study-build parser, a participant-facing assistant and a tool calculating an endpoint each need a different test plan. FDA's AI draft, ICH E6(R3) Section 4.3.4 and the risk-based CSA principles inform the proposed assessment [1] [2] [14].
3.1 The Three-Question Screen
FDA's draft guidance (FDA-2024-D-4689) conceptualizes risk through two fundamental dimensions: Model Influence and Decision Consequence [1]. In our operational framework adapting these concepts, we define three levels for each factor:
- Model Influence:
- Low Influence: The AI provides purely preparatory, informational, or drafting assistance subject to comprehensive, independent human review and verification before any action is taken.
- Medium Influence: The AI generates structured draft configurations, classifications, or operational flags that directly guide or prioritize human review, where human reviewers verify outputs against established objective criteria.
- High Influence: The decision substantially depends on the model output, with limited independent checking of that output before it is used. Rate dependence separately from the harm or disruption that an error could cause.
- Decision Consequence:
- Low Consequence: Minor operational or administrative inefficiency; no impact on participant safety, rights, protocol adherence, or clinical trial evidence integrity.
- Medium Consequence: Potential distortions in operational resource allocation, non-critical GCP documentation, or delayed query resolution, easily detected and corrected during routine data management.
- High Consequence: Direct threats to participant safety, unblinding of randomized treatments, compromise of primary or key secondary efficacy endpoints, or systematic bias in regulatory submissions.
Apply the three-question screen to every AI capability under review.
| Screening Question | Evaluation Criteria | Resulting Determination |
|---|---|---|
| Question 1: Unconfirmed writes | Can the function write to a regulated trial record without an authenticated, attributable human decision? | A function with unconfirmed writes is eligible only under the Tier 4 authorization below. For a function without unconfirmed writes, proceed to Question 2. |
| Question 2: Proximity to participants and decisions | What does the function do in the intended workflow? | 2A: Assign Tier 4 if the function determines endpoint value, severity grade, dose or eligibility without an independent human decision. Send advisory flagging to Question 3. 2B: Apply the Tier 1 navigation-only safeguards to participant-facing technical help. Route other participant-facing functions through the applicable decision or data-quality branch and evaluate the participant protections separately. 2C: Assign Tier 3 to advisory participant source-data quality checks. 2D: For other uses, go to Question 3. |
| Question 3: Model influence and decision consequence | Rate both using the definitions above. | Read the resulting tier from the lookup matrix. |
For Question 1, EMA §6.1.4 expects investigator authorization for sponsor changes to data entered by investigators or participants, whether automatic or manual [19]. Under the proposed Tier 4 exception, specify the function and controls in the approved protocol or data management plan, obtain written sponsor QA approval and investigator authorization for changes to investigator- or participant-entered data, and secure any applicable regulatory authorization for the intended use. Apply Questions 1 and 2 first. Where either assigns a tier or holds the function, that decision takes precedence. Only the remaining functions use the Question 3 matrix.
| Model Influence \ Decision Consequence | Low Consequence | Medium Consequence | High Consequence |
|---|---|---|---|
| Low Influence | Tier 1 (Informational Assistance) | Tier 2 (Draft Configuration) | Tier 2 (Draft Document Generation) |
| Medium Influence | Tier 1 (Workflow Prioritization) | Tier 2 (Operational Setup) | Tier 3 (Extraction & RBQM Surveillance) |
| High Influence | Tier 2 (Routing internal, noncritical training inquiries outside trial records) | Tier 3 (Prioritizing a noncritical operational review queue chiefly from a model score) | Tier 4 (High dependence and high consequence after Questions 1 and 2) |
Three-Question Screen for Clinical Trial AI Functions
Start with whether the tool can write or decide independently. For advisory functions use the influence and consequence matrix.
Scroll sideways for the full figure.
| Workflow | Decision route | Control set |
|---|---|---|
| Independent write | Reject unless the function is expressly qualified and authorized under Tier 4 | Tier 4 if authorized |
| Independent endpoint, severity, dose or eligibility determination | Use Branch 2A | Tier 4 |
| Technical navigation for participants | Branch 2B outside the matrix. Apply Pattern 1 and participant safeguards. | Tier 1 with navigation safeguards |
| Case A: coding suggestion queue | Question 3: medium influence, high consequence | Tier 3 |
| Case B: independently verified draft CSR narratives | Question 3: low influence, high consequence | Tier 2 |
| Advisory screening flags for possible eligibility | Question 3: medium influence, high consequence | Tier 3 |
| Query draft supplementing programmed edit checks | Question 3: medium influence, medium consequence | Tier 2 |
3.2 Four Operational Qualification Tiers
Synthesizing FDA's two-factor model risk, ICH E6(R3) Section 4.3.4 validation expectations, and GAMP 5 principles, EClinCloud recommends structuring eClinical AI governance into four operational tiers. Each tier defines specific testing postures and regulatory boundaries:
Tier 1: Informational and Administrative Assistance
- Scope: Typical matrix cell: low influence, low consequence. Staff-facing workflow assistance (e.g., software navigation bots answering how-to questions about EDC interfaces), clinical trial management system (CTMS) monitoring report proofreading, and administrative correspondence drafting.
- Participant-facing navigation safeguards. For participant-facing navigation assistants, apply Branch 2B outside the matrix. Lock the model, prompts and retrieval sources under Pattern 1 (Section 4.2). Restrict the assistant to technical navigation. Test refusal in all study languages before release and after prompt changes. Design the assistant to refuse medical advice, dosing questions and interpretation of instrument items. Use a fixed response directing health concerns to site staff or emergency services as appropriate, route the flagged exchange to site staff, and schedule periodic log review by site or safety staff for unreported adverse events. Refusal controls can miss out-of-scope answers, which is why the log review is part of the control set [21]. Submit participant-facing materials for applicable IRB/IEC review. See Section 6.2 for provider disclosure.
- Suggested Controls: Apply CSA unscripted and exploratory testing [14]; enforce zero direct write access to clinical databases; execute regression testing after SOP or software documentation updates; and maintain user interface disclaimers clarifying that advice is computer-generated.
Tier 2: Pre-Production Draft Configuration and Operational Classification
- Scope: Typical matrix cell: medium influence, medium consequence. Large language model parsing of clinical trial protocols into draft electronic Case Report Form (eCRF) schedules, visit matrices, and edit check specifications; AI-assisted setup of RTSM supply parameters and packaging configurations; automated classification of Trial Master File documents against standardized structures (e.g., TMF Reference Model); and automated draft translation of study administrative documents.
- Suggested Controls: Keep configuration tools upstream of production release. In-study drafting can also qualify where review occurs before use. Generated configurations are treated as unverified drafts. Data managers, biostatisticians, TMF specialists, and randomization leads independently review, test, and approve all configurations. Use scripted user acceptance tests for critical study configurations before release. ICH E6(R3) Sections 4.3.4(c) and (e) recommend validation before use, including standard functionality, study-specific configurations and interfaces [2].
Tier 3: Regulated Data Extraction, Capture QA, and RBQM Surveillance
- Scope: Typical matrix cell: medium influence, high consequence. Retrieval-augmented optical character recognition (OCR) extracting lab values from hospital records into eCRF drafts; real-time camera image quality checks in eCOA (detecting blur, glare, or incorrect angles in photographic diaries); automated personally identifiable information (PII) masking on source scans; and machine learning site-level risk scoring and Quality Tolerance Limit (QTL) surveillance within Risk-Based Quality Management (RBQM) platforms [25][29]. In the site-risk example, scoring has medium influence when staff make the final decision and high consequence when missing a signal could delay action on a serious data integrity or participant safety issue.
- Suggested Controls: Enforce approved model snapshots or bounded recalibration under Pattern 2 where appropriate; execute scripted UAT using golden benchmark test suites; implement source lineage architectures linking redacted working copies to investigator-controlled source records; record immutable sidecar inference and human disposition logs; and require authenticated human confirmation before the write. Assess the presentation of proposed values against EMA section 4.7 and the CDE auto-fill guidance, both described in Section 5.1 of this guide.
Tier 4: Highest-Control Clinical Functions
- Scope: Typical matrix cell: high influence, high consequence. Automated computer-vision quantification of clinical trial endpoints (e.g., echocardiographic LVEF scoring, radiographic lesion measurement, or digital histology) [27]; algorithmic digital biomarkers derived from wearable sensors supporting primary or secondary efficacy claims; and adaptive patient triage algorithms determining trial eligibility or investigational dosing. (Note: Human-led Independent Review Committee (IRC) workflows—where clinical specialists review medical images and record reader grades—remain human clinical evaluations rather than automated AI scoring tools.)
- Regulatory and Testing Posture: Tier 4 names a proposed control set. Device status under FD&C Act section 201(h), the MDR or the IVDR is a separate determination. In the United States, qualifying investigational device functions are governed by 21 CFR Part 812 (IDE) [13]. For investigations conducted under an approved IDE, mid-trial modifications must be evaluated under 21 CFR 812.35. In the European Union, clinical investigations of software that is a medical device fall under MDR Articles 62 through 82 and Annex XV, typically conducted under ISO 14155 [30]. Software that qualifies as an in vitro diagnostic device falls under the performance-study requirements of Regulation (EU) 2017/746 (IVDR). Digital histology software analyzing tissue specimens can fall in this group, depending on intended purpose. Assess EU AI Act classification separately (see Section 6.2). EMA expects pivotal hypothesis-testing models and preprocessing to be frozen and documented before database lock and supports prospective clinical validation in high-impact settings [9]. EClinCloud also suggests blinded independent adjudication where suitable.
Context-of-Use Risk Classification Matrix for Clinical Trial Software Functions
Evaluating trial software functions by model influence and decision consequence establishes proportionate qualification controls from exploratory testing to prospective clinical validation.
Scroll sideways for the full figure.
| Qualification Tier | Typical matrix cell or assigned branch | Representative Software Functions | Regulatory Framework & Legal Scope | Suggested Testing & Governance Controls |
|---|---|---|---|---|
| Tier 1: Informational & Administrative Assistance | Low influence, low consequence | Staff-facing navigation bots; CTMS monitoring report proofreading; administrative drafting. | EClinCloud reading: the FDA draft exclusion depends on effects on patient safety, drug quality or study-result reliability. Article 50(1) provider duty where the EU Regulation applies. | CSA unscripted and exploratory testing; zero database write permissions; interface disclaimers; regression checks on SOP updates. |
| Tier 1 (Participant-Facing Exception) | Branch 2B outside the matrix | eCOA patient mobile assistants answering device charging, alarm scheduling, or navigation questions. | Article 50(1) provider duty where the Regulation applies. | Pattern 1 lock. Test refusal in all study languages. Direct health concerns to site staff or emergency services, route flags and review logs. Submit applicable materials for IRB/IEC review. |
| Tier 2: Pre-Production Draft Configuration | Medium influence, medium consequence | Draft eCRF/visit matrices, RTSM rules and eTMF classification. Independently verified draft CSR safety narratives are low influence, high consequence (Case B). | EClinCloud reading: FDA draft scope depends on effects on safety, drug quality and study-result reliability. Independent review informs the assessment. | ICH E6(R3) 4.3.4(c) and (e) recommend validation before use. EClinCloud suggests scripted UAT and review of drafts. |
| Tier 3: Regulated Data Extraction & Surveillance | Medium influence, high consequence | Lab OCR extraction into eCRF drafts; eCOA real-time image quality checks; automated PII redaction; RBQM site risk scoring and QTL monitoring. | Part 11 applies where the system is in scope. ICH E6(R3) 4.2.2 recommends audit-trail controls. | Apply protocol locks for extraction or bounded recalibration for qualified RBQM scoring. Test using golden cases and retain sidecar logs. Assess how proposed values are shown against EMA section 4.7 and the CDE auto-fill guidance, both described in Section 5.1 of the article. |
| Tier 4: Highest-Control Clinical Functions | High influence, high consequence | Automated echocardiographic LVEF scoring; radiographic lesion quantification; wearable sensor digital biomarkers supporting primary efficacy claims. | Evaluate device status separately. Where the function meets the device definition, apply the relevant FDA investigational or EU MDR/IVDR requirements and assess AI Act scope. | EMA expects models and preprocessing for hypothesis testing frozen and documented before database lock. EClinCloud suggests blinded adjudication where suitable. |
Tiers and suggested controls represent EClinCloud recommendations; legal classifications reflect applicable statutory frameworks.
3.3 Worked Classification Examples for Common Trial Software Functions
To demonstrate how the three-question screen applies to common trial software functions, consider five representative implementations:
Case A: MedDRA Coding Suggestions in EDC: An EDC platform integrates an embedding model that suggests Medical Dictionary for Regulatory Activities (MedDRA) Lowest Level Terms and Preferred Terms for verbatim adverse event terms entered by site coordinators.
Screening: The EMA guideline names coding in its scope [19]. Auto-commit uses the Question 1 authorization rule. In the suggestion queue a medical coder actively accepts or overrides the code. Branch 2D sends it to Question 3. Influence is medium because the suggestion directs review. Consequence is high because erroneous coding can obscure safety signals.
Result: Tier 3 (Regulated Data Extraction & Coding Assistance). Operating inside active clinical trial data management, it requires Tier 3 controls, including golden-set dictionary validation, override logging, and mandatory human medical coder confirmation.
Case B: Automated Clinical Study Report (CSR) Safety Narrative Generation: A CDMS or medical writing tool uses an LLM to draft patient safety narratives from adverse event databases, lab listings, and concomitant medications.
Screening: The output remains a draft. Under Question 3 the consequence is high and influence low when a medical writer independently verifies the content against the source. EClinCloud considers the FDA draft's Section II scope test to depend on the actual influence on patient safety, drug quality or study result reliability. Independent verification supports this assessment without deciding the legal scope [1].
Result: Tier 2. Validate the draft workflow and record the review against source listings.
Case C: eCOA Participant Navigation Assistant: A conversational assistant embedded in an eDiary mobile application answers participant questions regarding device charging, diary alarm scheduling, and app navigation.
Screening: Question 1: Write privilege? No write access to survey responses. Question 2: Proximity? Directly participant-facing, triggering Branch 2B.
Result: Tier 1 with navigation safeguards. Apply the Pattern 1 lock and participant safeguards in Section 3.2.
Case D: Automated Patient Eligibility Pre-Screening Tool: An electronic health record (EHR) screening algorithm flags potential trial candidates for site study coordinators based on protocol inclusion/exclusion criteria.
Screening: The EMA guideline names recruitment and eligibility [19]. Site staff see an advisory flag and independently decide eligibility. Under Question 3 influence is medium and consequence high.
Result: Tier 3. Test a sample of unflagged as well as flagged records. An algorithm that determines eligibility without an independent human decision uses Branch 2A and the Tier 4 controls.
Case E: Automated Discrepancy / Query Drafting Assistant: A data management tool analyzes eCRF data against protocol rules and drafts proposed queries for data manager review before sending to sites.
Screening: The EMA guideline names data clarification and query processes [19]. The tool supplements programmed edit checks. The data manager verifies proposed queries before transmission. In this example it prioritizes review, making influence medium and consequence medium for delayed resolution.
Result: Tier 2. Test draft accuracy and inspect a sample of unflagged records for missed discrepancies. Reassess the consequence if it replaces an established edit check.
4. Version Control, Change Governance, and Guardrails Against Silent Drift
Conventional computerized system validation tests a controlled software version, configuration, system state and operating environment. AI deployments can change through provider checkpoints or parameters, updates to retrieval sources, and shifts in the inputs encountered across regions or clinical settings. Record which inputs and runtime conditions the validation covers.
4.1 Deployment Realities: Self-Hosted Digests vs. Hosted Cloud Models
A practical engineering distinction that quality assurance leads should recognize is that version-locking mechanisms differ between self-hosted models and third-party commercial cloud APIs:
- Self-hosted and open-weight models. Record SHA-256 fingerprints of the approved weights, container images, configuration and inference dependencies. Restrict deployments to approved artifacts and verify the fingerprints of what the runtime actually loads. Change restrictions and release checks provide the control. A fingerprint identifies the version being checked.
- Commercial Cloud and Hyperscaler APIs: When trial software connects to proprietary cloud models via external APIs, providers generally do not expose model weights, neural architectures, or cryptographic digests. Furthermore, hyperscalers retire model checkpoints on commercial schedules that can conflict with multi-year clinical trial timelines. In addition, temperature settings of zero reduce but do not completely eliminate output non-determinism across distributed GPU clusters. For cloud APIs, ask for the provider’s available evidence of change control and a pinned, dated model identifier (e.g., referencing a specific static release date rather than a floating alias such as
-latest).
Guardrails Against Silent Drift for Hosted APIs: A stable API identifier confirms the label. Test results reveal whether behavior has changed. To guard against silent parameter adjustments under an unchanged identifier, we recommend that sponsors mandate scheduled periodic golden-set re-runs (e.g., monthly or quarterly). The vendor or sponsor executes an automated benchmark test suite through the active API endpoint, measuring the share of golden-set numeric fields that differ from the reference, including missing fields, against a predefined threshold. Hold use of the affected function for investigation and retest if the share exceeds the threshold.
To manage third-party API lifecycles in multi-year studies, we suggest that sponsors incorporate a Forced-Migration and Deprecation Protocol into vendor Quality Agreements:
- Contractual Advance Notice: The vendor contractually agrees to provide advance written notice prior to the retirement, deprecation, or parameter modification of any underlying third-party model endpoint (an illustrative starting target of 30 days, tied to the provider's published commercial notice policy).
- Golden-Set Concordance Testing: Upon receiving deprecation notice, the vendor and sponsor execute comparative evaluation comparing the retiring model against the designated replacement across a validated golden benchmark test suite, measuring output concordance against pre-defined tolerance bands.
- Impact Assessment Under ICH E6(R3): In accordance with ICH E6(R3) Section 4.3.4(c), which states that validation should assess the impact of software changes across both historical and future trial data, data management and biostatistics teams document whether model migration could introduce systematic bias or shifts in data extraction patterns [2].
- Documented TMF Bridging Report: The change control rationale, comparative evaluation results, and risk assessment are documented in a formal Trial Master File bridging report approved by QA prior to production cutover.
4.2 Three Change-Governance Patterns and EMA's Position on Frozen Models
To align software maintenance with regulatory expectations, we recommend assigning every AI capability to one of three change-governance patterns prior to First Patient In (FPI):
| Change Governance Pattern | Applicable Scope & Tiers | Technical Mechanism | Quality & Validation Governance |
|---|---|---|---|
| Pattern 1: Protocol-Locked Immutable Snapshot | Tier 2 study-build and in-study drafting tools, Tier 3 regulated extraction, Tier 4 algorithms frozen for the study under the applicable EMA position, and Tier 1 participant navigation. | Freeze model weights SHA-256 (or pinned dated API identifier), system prompts, decoding parameters, and retrieval indices per study build. | Zero mid-trial changes permitted without formal change control, regression testing against golden benchmark suites within defined tolerance bands, and QA approval. |
| Pattern 2: Bounded Operational Recalibration | Tier 3 RBQM surveillance / site risk scoring models. | Recalibration occurs strictly within pre-defined operational boundaries, with documented convergence criteria and automated rollback triggers (where technically feasible on self-hosted or supported versions). | Bounded change protocol pre-specified in the trial Data Management Plan or Monitoring Plan; assess changes as ICH E6(R3) Section 4.3.4(c) recommends [2]. |
| Pattern 3: Controlled Tenant Release with Zero Write Access | Tier 1 staff-facing workflow assistants and administrative tools. | Regular updates under vendor software development lifecycles and CSA unscripted/exploratory testing. | Logical isolation from clinical databases; zero write permissions; regression testing on updated SOPs and prompt guardrails. |
EMA's Position on Frozen Models in Pivotal Trials: For late-stage pivotal trials, Section 2.3.3.3 of the EMA reflection paper calls for prospective performance testing before deployment in high regulatory impact settings, such as primary-endpoint analysis. It expects the preprocessing pipeline and models used for hypothesis testing to be frozen and documented before database lock and subsequent unblinding. EMA rejects incremental learning in this context and expects regulatory interaction to amend the statistical analysis plan if a model changes during the trial [9]. Use Pattern 1 for confirmatory analysis. Evaluate Pattern 2 separately for an operational use such as RBQM site scoring, including whether the change could alter endpoint data capture or monitoring.
4.3 Staged Pre-Release Validation and Adversarial UAT Assertions
Drawing inspiration from the staged environments piloted in the UK MHRA AI Airlock program [21], EClinCloud recommends that sponsors implement a three-stage pre-release testing pipeline prior to deploying Tier 3 or Tier 4 software updates:
- Simulation Stage: Conduct multidisciplinary stress-testing workshops bringing together data managers, software engineers, and clinical coordinators to evaluate model behavior under challenging simulated conditions (such as low-resolution smartphone photos, multilingual clinical records, and extreme clinical values) [21].
- Virtual Stage: Replay frozen historical trial datasets through the updated algorithm within a secure, isolated sandbox environment to evaluate concordance against validated historical baselines, verifying that output variance remains strictly within pre-defined tolerance bands [21].
- Shadow-Deployment Stage (EClinCloud Operational Extension): As an operational extension beyond what the MHRA Phase 2 sandbox evaluated (which tested retrospective and synthetic data separate from clinical pathways) [21], sponsors may run the updated model silently in parallel with the existing validated production system for an illustrative duration of 14 to 30 days. Before running in shadow mode, confirm that the processing stays within the protocol, consent and vendor data-processing terms. Do not show shadow outputs to site staff or write them to study records. This allows engineers to quantify output concordance across live clinical data streams before switching primary workflows.
During formal User Acceptance Testing (UAT), sponsor QA teams should execute targeted adversarial assertions mapped to the relevant tiers:
- Runtime Provenance and Metadata Verification (Tiers 2, 3, 4): Query the inference API and assert that every transaction returns an unambiguous
model_id_and_version(cryptographic SHA-256 digest for self-hosted models; pinned dated release identifier for cloud APIs), andprompt_or_config_hashmatching the approved vendor qualification dossier. - Golden Perturbation and Locale Matrix (Tiers 3, 4): Execute an automated benchmark pack containing corrupted clinical scans, skewed mobile photos, bilingual medication labels, and regional numeric variations (such as European decimal commas vs. American decimal points, and
DD-MM-YYYYvs.MM-DD-YYYYdate formats). Measure the share of benchmark numeric fields differing from the reference, including omitted fields. Compare it with the predefined threshold and hold release for investigation and rerun if it exceeds the threshold. - Refusal and Uncertainty Calibration (Tiers 1 Participant-Facing, 3, 4): Submit completely illegible scans, cropped documents, and out-of-scope medical questions. Assert that the system triggers an explicit low-confidence refusal flag, prompting manual human entry or coordinator escalation, rather than fabricating plausible values.
- Uncommitted-Draft Isolation (Tiers 2, 3): Attempt programmatic API and front-end interface submissions to transition an eTMF document to Filed status or an eCRF data field to Saved status directly from an AI output without an authenticated human review event. Assert that the computerized system strictly denies write access, enforcing human confirmation.
5. Recording Inference and Disposition for Inspection
For closed systems subject to Part 11, 21 CFR 11.10(e) requires secure, computer-generated, time-stamped audit trails that independently record the date and time of operator actions that create, modify or delete electronic records [31]. Complementing this binding rule, FDA Electronic Systems Q&A Q12–Q14 [12], ICH E6(R3) Section 4.2.2 [2], EMA Guideline EMA/INS/GCP/112288/2023 [19], and China's CDE computerized-systems guideline [6] recommend that computerized systems maintain complete chronological traceability. To prepare for inspection, test whether the history can be reconstructed: Can the sponsor reconstruct the original data, the inference event and its subsequent disposition, including the responsible person or authorized machine actor and the resulting record state?
5.1 Forensic Failure Modes: Ghost Pre-Fills and Working-Copy Source Lineage
When software vendors treat artificial intelligence as a transient front-end user convenience rather than an inspectable computerized process, two serious forensic compliance failures emerge:
- Ghost Pre-Fills: If an optical character recognition model or language model automatically populates an eCRF field in a web browser, and a clinical research coordinator simply clicks Save, standard database audit trails log only that the coordinator entered the data. A save event alone leaves an auditor unable to distinguish an automated suggestion from a value the coordinator verified. Test whether other retained evidence reconstructs the workflow. EMA section 4.7 advises against prepopulating or automatically filling fields unless they are non-editable and derived from previously entered data [19]. China’s CDE guideline generally advises against setting data points to auto-fill unless they are non-editable or derived from previously entered data (printed pages 3–4) [6]. Both address auto-fill without naming AI. EClinCloud reads model output written into an editable field as prepopulation under both. During qualification, check whether the interface keeps the suggestion separate from the field until a person accepts or enters it. For editable pre-fills, document the risk and the applicable guidance assessment. Treat a pre-fill outside the applicable conditions as a design issue to resolve before release. ICH E6(R3) Section 4.2.2 recommends maintaining relevant workflow actions for higher-criticality data [2]. FDA and EMA’s good AI practice principles support assessing the human review step in the context of use [11].
- Broken Source Lineage After Automated Privacy Redaction: When clinical software applies computer-vision or natural language processing filters to redact patient names, national identity numbers, or facial features from uploaded medical records or eCOA photographic diaries, vendors sometimes overwrite or discard the original unredacted file. This practice undermines data integrity. FDA’s electronic systems Q&A (Q3) describes a certified copy as one verified to have the same information, including context, content and structure, as the original [12]. A masked or redacted document is therefore not a certified copy; it is a redacted working copy. Original source documents containing patient identifiers should remain under the legal and physical control of the clinical investigator at the investigative site (or within an investigator-controlled, access-segregated partition), adhering to ICH E6(R3) Principle 1.6 and Section 3.16.1(t) regarding participant confidentiality and applicable privacy laws [2]. In our recommended architecture, the sponsor-accessible system retains an encrypted, redacted working copy and links it to the site-held original using a source pointer and a site-generated SHA-256 digest, establishing verifiable source lineage to support source data verification during audits.
5.2 A Suggested Sidecar Inference-and-Disposition Event Record
EClinCloud proposes an immutable inference event for each Tier 2, Tier 3 or Tier 4 invocation. Append a distinct disposition linked to the inference UUID. For a deferred suggestion, preserve the initial state and append any later disposition, retaining both events in order.
inference_event_uuid: A universally unique identifier establishing direct relational linkage between the AI event and the corresponding clinical database record (e.g., eCRF field, eCOA upload, or eTMF artifact).model_id_and_version: Pinned model identifier, container digest, or static cloud release date. FDA’s DHT guidance recommends recording when and how each DHT was updated. EClinCloud applies that recordkeeping approach to each model record [13].prompt_or_config_hash: SHA-256 cryptographic digest of system instructions, prompt templates, RAG vector database snapshot, temperature settings, and decision threshold parameters.source_pointer_and_hash: Unique locator and SHA-256 hash linking the encrypted, redacted working copy to the original. Compute the original’s SHA-256 at the site when making the working copy and store it with the copy. Verify it by site-side recomputation.raw_ai_output_and_score: The unedited machine suggestion and its confidence score where available, retained independently of what happens to the record.disposition_state: RecordHUMAN_REVIEWED,AUTHORIZED_AUTOMATEDorUNCOMMITTED. A deferred item remains uncommitted until a later disposition permits a write.human_disposition_action: For actual human review, recordACCEPTED_UNCHANGED,EDITED_AND_CONFIRMED,REJECTED_MANUAL_ENTRYorDEFERRED. Leave this field empty when no review occurred.machine_actor_and_authorization_ref: For an authorized automated disposition, identify the executing service or job and link the approved context and applicable authorizations from Section 3.1. An automated write without those records fails the test.final_committed_value: For a committed human-reviewed or authorized automated disposition, retain the exact result written to the clinical record. Leave this field empty on an uncommitted disposition.reviewer_id_and_role: When human review occurs, record the authenticated identity and role, supporting role segregation recommended in EMA Guideline 112288/2023 [19].override_reason_code: When human review occurs, record the reason whenever the reviewer overrides a high-confidence suggestion or accepts a low-confidence prediction, supporting quality root-cause analysis.inference_and_disposition_timestamps: High-precision UTC timestamps for execution and each subsequent disposition, including human sign-off where it occurred, consistent with FDA Q14 system clock recommendations and enabling review dwell-time analytics [12].review_dwell_seconds: For completed human review, record the duration from presentation to submission, providing operational metadata to evaluate review rigor.
Suggested Inference and Disposition Records and Verification
Record the inference and append its attributable disposition. Use separate evidence for human review, authorized automated writes and outcomes that remain uncommitted.
Scroll sideways for the full figure.
| Field Name | Data Type & Example Format | Regulatory & Governance Rationale | Recommended UAT Verification Method |
|---|---|---|---|
| inference_event_uuid | UUID string (e.g., 'a1b2c3d4-...') | Relational foreign key linking AI inference directly to clinical database record (eCRF, eCOA, eTMF). | Verify the event resolves to its intended record or draft and all attempts remain traceable. |
| model_id_and_version | String: Model ID + SHA-256 / Dated ID | Captures the exact model version and deployment identifier. FDA’s DHT guidance recommends an update record for each DHT. EClinCloud extends that to model records. | Assert the API returns model_id_and_version matching the approved dossier. |
| prompt_or_config_hash | String: 64-char SHA-256 hex digest | Identifies the approved snapshot of prompts, decoding parameters and retrieval indices. | Recalculate SHA-256 over prompt and config payload; assert match with approved release hash. |
| source_pointer_and_hash | Source locator and SHA-256 computed at the site when the working copy is created | Establishes source lineage linking redacted working copy to site-held original record. | Recompute at the site and compare with the retained hash. |
| raw_ai_output_and_score | JSON object: text/value + float score | Persists unedited machine suggestion and confidence score to reconstruct initial prediction. | Assert output payload contains original prediction prior to human interface display. |
| disposition_state | HUMAN_REVIEWED / AUTHORIZED_AUTOMATED / UNCOMMITTED | Distinguishes the outcome from the inference event. | Replay a deferred suggestion followed by a decision. Verify both dispositions remain linked and ordered. |
| machine_actor_and_authorization_ref | Executing service or job and approval record ID for authorized automation | Identifies machine responsibility and Section 3.1 authorization. | Reject a seeded automated write without applicable approval. Verify actor and approval reference on an allowed write. |
| human_disposition_action | Conditional on actual human review: ACCEPTED_UNCHANGED, EDITED_AND_CONFIRMED, REJECTED_MANUAL_ENTRY or DEFERRED. | Aligns with FDA-EMA Guiding Principles 1 & 8 to evaluate human-AI interaction and oversight. | Exercise the human actions and verify the field is empty when no person reviewed the output. |
| final_committed_value | String / Numeric / Coded entry | Captures the value written by human review or authorized automation. | Compare the written value with the disposition. Verify that a deferred or uncommitted event has no committed value. |
| reviewer_id_and_role | String: Authenticated user ID + role | Records human review when it occurs. | Compare identity and role with the review session. Verify no human identity is recorded for an automated-only disposition. |
| override_reason_code | Enum / Categorical code + optional text | Mandatory justification when human overrides high-confidence AI or accepts low-confidence output. | Attempt submission without reason code on override; assert interface validation blocks commit. |
| inference_and_disposition_timestamps | ISO 8601 UTC string (YYYY-MM-DDThh:mm:ssZ) | Records UTC inference and disposition times, including any actual human sign-off. | Assert timestamps are captured from synchronized NTP system clocks in UTC format. |
| review_dwell_seconds | Float / Integer (seconds) | Optional RBQM input for completed human review. | Check elapsed time from presentation to submission. Leave empty if no human review occurred. |
All fields, reason codes, and verification assertions represent EClinCloud suggested implementations.
5.3 Leveraging Audit Metadata for RBQM Surveillance and Inspection Readiness
ICH E6(R3) Section 4.2.3 recommends that sponsors conduct risk-based reviews of clinical data and metadata, including audit trails [2]. By aggregating transaction-level audit metadata across investigative sites, central data managers and Risk-Based Quality Management (RBQM) leads can identify operational anomalies:
- Monitoring Automation Complacency via Review Dwell Time: Drawing on UK MHRA AI Airlock Insight 2 regarding the potential erosion of human oversight over time [21], central monitoring dashboards can track the
AI Acceptance RatealongsideMean Review Dwell Time. If investigative site staff consistently recordACCEPTED_UNCHANGEDdispositions on complex multi-page documents with dwell times below an illustrative threshold (e.g., under two seconds per page), this operational signal serves as a prompt for inquiry. Data management leads can investigate whether coordinators are verifying source documents or accepting outputs superficially, scheduling targeted retraining where appropriate. However, QA leads must recognize that dwell time is an operational proxy with inherent limits: a low dwell time alone is an ambiguous signal (for instance, a coordinator may have thoroughly inspected paper source records prior to opening the software screen), and a high dwell time can reflect a screen left open. EClinCloud recommends reporting dwell time at site level unless site agreements and applicable data-protection and employment rules permit individual reporting. - Tracking Regional Distribution Shifts via Override Rates: Monitoring the
Human Override Rateacross trial sites, countries, and laboratory networks alerts clinical operations to emerging distribution shifts. A sudden increase in override rates at newly onboarded international sites can serve as a prompt to investigate differences in local medical abbreviations, hospital record layouts, or laboratory units, identifying the need for localized UAT adjustments. EClinCloud suggests calculating both rates by function, site and period. Count only completed human-reviewed dispositions. Divide acceptances without edit by that denominator for the acceptance rate and divide edits plus rejections by the same denominator for the override rate. Exclude authorized automated events. Report deferrals separately and include a deferred item in the rates only after its completed human disposition. Show the denominator with each rate. - Exporting Chronological Audit Trails: For closed systems within Part 11 scope, 21 CFR 11.10(b) addresses inspection copies and 11.10(e) addresses audit trails [31]. ICH E6(R3) Section 4.2.2(c) recommends interpretable audit trails and logs for review [2]. FDA Q12 recommends retaining audit trails in a searchable and sortable format and accepts a static format such as PDF where that is impractical [12]. EMA §6.2.1 addresses review and export [19]. For vendor qualification, EClinCloud suggests verifying export of the complete audit log in a documented machine-readable format, such as CSV or XML, that opens without proprietary vendor tools. Obtain a corresponding human-readable presentation, such as a PDF. Verify both show the inference UUID, model and configuration identifiers, source pointer, raw suggestion, disposition state, committed value when present and synchronized UTC timestamps. For a human decision include the reviewer and applicable reason. For authorized automation include the machine actor and authorization reference.
6. The EU AI Act: Where Clinical Trial Software Actually Sits
Read together, the AI Act and Regulation (EU) 2026/1744 set a tiered structure. Start with scope, then determine classification and any transparency obligation [7] [8].
6.1 The Scientific Research Exclusion: Article 2(6), Recital 25, and Article 2(8)
The threshold legal question for any clinical trial tool is whether it falls within the research and development carve-out established in Article 2(6) [7]. Article 2(6) provides verbatim:
"This Regulation does not apply to AI systems or AI models, including their output, specifically developed and put into service for the sole purpose of scientific research and development." [7]
Recital 25 explains this exclusion, stating that AI systems developed and put into service solely for scientific research and development should be excluded from the scope of the Regulation, while noting that "any other AI system that may be used for the conduct of any research and development activity should remain subject to the provisions of this Regulation." [7]
Article 2(6) asks why the AI system was developed and put into service. For a commercially supplied eClinical tool, document the specific intended purpose and deployment. Stakeholders asked the Commission for clearer guidance on applying the research exclusions in areas including medicines and medical devices [7] [32]. The following two readings are editorial interpretations to test against the facts of each deployment.
- Research-specific deployment. A protocol parser developed and put into service for one defined scientific research purpose could present a stronger case under Article 2(6). Capture the design brief, contract, permitted users and deployment constraints that demonstrate that purpose. A clinical-trial label gives limited evidence on its own [7].
- Reusable operational service. A platform that serves multiple purposes, such as study workflows and general document processing, presents a weaker case. Document the provider and deployer roles using the actual development, supply and deployment arrangements. Recital 25 distinguishes AI systems developed and put into service for the sole purpose of scientific research from other systems that happen to be used in research [7].
Qualification screen. Use these three cases to identify evidence to take to the applicable legal and quality reviewers. Record the classification and reconsider it if the intended purpose or deployment changes.
- General-Purpose and Multi-Use Software: Under Recital 25, other AI systems used for research remain subject to the Regulation. In our reading that covers general foundation models, commercial support chatbots and general enterprise OCR tools used in research [7].
- Commercially Licensed Trial-Specific Software: For an EDC lab OCR tool or an eTMF classifier within the Act’s scope, determine whether its intended purpose meets an Annex I or Annex III classification and whether an Article 50 transparency duty applies [7].
- Pre-market research, testing and development. Article 2(8) excludes research, testing and development regarding AI systems or models before market placement or putting into service, subject to applicable Union law. Its exclusion does not cover testing in real-world conditions. That is distinct from Article 2(6). Nonbinding MDCG 2025-6 Question 25 treats high-risk medical-device AI clinical investigations and performance studies as real-world testing and explains that Article 60(1) preserves applicable Union or national law, including the MDR and IVDR [7] [30] [33] [34].
6.2 Digital Omnibus Amendments: Timelines, Articles 4a, 6(1a)–(1b), 6(3), and 50
Regulation (EU) 2026/1744 (the Digital Omnibus on AI, published in the Official Journal on July 24, 2026, and entered into force July 27, 2026) enacted key amendments to Regulation (EU) 2024/1689 [8]:
Application dates for Chapter III Sections 1–3. Article 113 as amended applies those three sections, with the exception of Article 6(5), from December 2, 2027 for high-risk systems classified under Article 6(2) and Annex III, and from August 2, 2028 for systems classified under Article 6(1) and Annex I [8]. For Article 6(1), check the product or safety-component criterion and the third-party conformity assessment requirement. Under Article 6(1c), a third-party assessment triggered solely by risks unrelated to health and safety fails to meet Article 6(1)(b) [8]. Article 50 has a separate timeline below. Article 2(13) allows specific obligations in Articles 9 to 15 and 17 to 25 to be limited for Article 6(1) high-risk systems. Annex I Section A legislation must provide equivalent or greater protection for health, safety or fundamental rights, and the limitation must preserve the Act’s overall level of protection. The Commission must adopt delegated acts by August 2, 2027 specifying the systems, obligations, conditions and scope [7][8].
Article 50(1) transparency from August 2, 2026. Where the Act applies, providers of AI systems intended to interact directly with people must inform them of the interaction, “unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use” [7]. Article 50(5) sets the deadline at the first interaction or exposure. For study qualification, EClinCloud recommends checking for a legible disclosure on participant- and site-facing conversational assistants.
Clarifications to Article 6 High-Risk Classification Rules: Regulation (EU) 2026/1744 added safety-component rules in Article 6(1a) and (1b) [8]. Under Article 6(1a), an AI system used solely for non-safety-related aspects of user assistance, performance optimisation, service efficiency, automation or convenience, or quality control falls outside the Act’s safety-component classification [8]. Conversely, Article 6(1b) clarifies that safety-component status applies wherever a failure or malfunction of the AI system would endanger health and safety [8]. In our analysis, most operational eClinical tools (such as CTMS schedulers or TMF indexing tools) sit entirely outside Annex I and Annex III [7]. For software that touches an Annex III domain, Article 6(3) establishes that an AI system shall not be considered high-risk if it does not pose a significant risk of harm to health, safety, or fundamental rights AND satisfies at least one of four specific conditions: (a) it performs a narrow procedural task; (b) it improves the result of a previously completed human activity; (c) its purpose is to detect decision-making patterns or deviations from prior patterns, with proper human review required if it would influence or replace a previously completed human assessment; or (d) it performs a preparatory task to an assessment [7]. An Annex III system that profiles natural persons remains high-risk under Article 6(3) [7]. Under Article 6(4), a provider claiming an Article 6(3) derogation must document its assessment before placing the system on the market or putting it into service and register the system in the EU database under Article 49(2), obligations that take effect on December 2, 2027 [7][8].
Conditional Bias Processing Under Article 4a. Article 4a applies from July 27, 2026 under Article 4 of the amending Regulation and establishes a strictly safeguarded legal exception for processing special categories of personal data (including health, genetic, and biometric data) for the purpose of detecting and correcting algorithmic bias [8]. Under Article 4a(1), providers of high-risk AI systems may process special categories of personal data strictly for bias detection and correction subject to six cumulative conditions in addition to the GDPR or other applicable EU data-protection law:
(a) bias detection and correction cannot be effectively fulfilled by processing other data, including synthetic or anonymised data;
(b) technical limitations on the re-use of personal data, and state-of-the-art security and privacy-preserving measures, including pseudonymisation;
(c) access strictly restricted to authorized personnel bound by confidentiality;
(d) personal data are not transmitted, transferred or otherwise accessed by other parties;
(e) personal data are deleted once bias has been corrected or at the end of the retention period, whichever comes first; and
(f) records of processing activities are maintained stating reasons why processing was strictly necessary [8].
Article 4a(2) extends the permission to providers and deployers of other AI systems and models, and to deployers of high-risk AI systems, where processing is strictly necessary to detect and correct biases likely to affect the health and safety of persons, have a negative impact on fundamental rights, or lead to discrimination prohibited under Union law, and where all paragraph 1 conditions are met. The provision states: "This paragraph does not create any obligation to conduct such bias detection and correction." [8]. In EClinCloud’s reading, this matters when bias testing of trial software, such as eligibility pre-screening or endpoint measurement, would use participants’ health data.
EU AI Act Applicability and Implementation Boundaries for Clinical Trial Software
Scope, high-risk classification and transparency depend on intended use. The dates for Chapter III Sections 1–3, except Article 6(5), are separate from Article 50.
Scroll sideways for the full figure.
| AI System Category / Workflow | EU AI Act Classification & Legal Basis | Effective Application Date | Suggested Sponsor & Vendor Qualification Actions |
|---|---|---|---|
| Standalone eClinical Workflow Software (EDC, CTMS, eTMF, RTSM) | EClinCloud reading: typically outside Annex I and Annex III. Confirm against the intended purpose. | Confirm scope and intended purpose. | Document intended purpose and evaluate Articles 2 and 6 for the actual deployment. |
| Trial-Specific AI Touching Annex III Domains with Art 6(3) Derogation | Non-High-Risk via Art 6(3) derogation (requires: no significant risk, one of four conditions, and no profiling). | December 2, 2027 for the applicable Chapter III Sections 1–3 rules, except Article 6(5). | Document the Article 6(3) assessment before market placement or entry into service. Prepare for EU database registration under Article 49(2) by December 2027. |
| AI-Enabled Medical Devices / SaMD under MDR 2017/745 or IVDR 2017/746 | Annex I high-risk AI system when Article 6(1) applies. Article 6(1c) excludes assessments triggered solely by risks unrelated to health and safety from Article 6(1)(b). | August 2, 2028 for Chapter III Sections 1–3, except Article 6(5), where Article 6(1) and Annex I apply. | For MDR devices, determine whether clinical investigation requirements in Articles 62–82 and Annex XV apply. Consider ISO 14155 in that clinical investigation context. For IVDR devices, assess the applicable performance-study requirements separately. See Sources 30 and 33. |
| Participant-Facing Conversational Assistants (technical help) | Article 50(1) provider disclosure duty where the Regulation applies, subject to its obvious-interaction exception. | August 2, 2026 | Verify provider has implemented legible, accessible interface disclosures informing participants they are interacting with AI; confirm IRB/IEC approval. |
| Special-Category Health Data Processing for Bias Mitigation (Article 4a) | Article 4a(1) covers high-risk AI providers under cumulative safeguards. Article 4a(2) conditionally extends the permission to high-risk AI deployers and other AI providers and deployers. | July 27, 2026 under Article 4 of Regulation (EU) 2026/1744. | Apply all conditions in Section 6.2 and applicable data-protection law. Under paragraph (2), establish strict necessity and the applicable health, safety, fundamental-rights or prohibited-discrimination boundary. The permission creates no general duty to perform bias processing. |
| AI Exclusively Developed and Put into Service for Scientific R&D | Article 2(6): sole purpose scientific research and development. Separately, Article 2(8) excludes pre-market research, testing and development, with real-world testing outside that exclusion. | Assess the distinct scope conditions. | Examine whether commercial SaaS qualifies under sole-purpose R&D test; do not assume automatic exemption for multi-tenant commercial platforms. |
Dates for Chapter III Sections 1–3 exclude Article 6(5). Check Article 2 research scope separately for the actual deployment.
7. An Illustrative Phased 90-Day Implementation Sequence for Sponsors and CROs
Transitioning from ad-hoc pilot scripts to an inspection-defensible AI governance framework requires an orderly, staged implementation plan. Sponsors and CROs can operationalize these recommendations through an illustrative 90-day sequence, whose actual timeline will depend on vendor software release cycles and internal change management procedures.
7.1 Days 1 to 30: Inventory, Context-of-Use Screening, and Regulatory Triage
- Comprehensive AI Asset Inventory: Catalog every artificial intelligence, machine learning, and natural language processing capability embedded across active and planned clinical trial systems, spanning eCRF protocol parsers, medical OCR, eCOA image quality scanners, eTMF indexers, RBQM anomaly detectors, and medical coding assistants.
- Context-of-use screening. Apply the three-question screen in Section 3.1 and record the evidence supporting each tier assignment.
- Legal status and medical device triage. Begin with Tier 4 and evaluate any function whose intended purpose may meet the device definition under FD&C Act Section 201(h), MDR or IVDR. Determine the applicable investigational route [13] [30] [33]. Check the disclosure provision in Section 6.2 for participant- or site-facing conversational assistants.
7.2 Days 31 to 60: SOP Re-Baselines, Quality Agreements, and Sidecar Audit Schema
- SOP updates. Distinguish (1) binding regulations, including applicable provisions in 21 CFR Parts 11, 50, 56, 312 and 812, the EU Clinical Trials Regulation, MDR, IVDR and AI Act, and China NMPA Announcement 2026 No. 50 [18] [30] [33] [7] [5], (2) adopted guidance, including EMA’s computerized-systems guideline and ICH E6(R3), a guideline Article 47 tells sponsors and investigators to take into account in EU trials [19] [2] [18], (3) agency recommendations, including the FDA eSystems Q&A, DHT, CSA and E6(R3) guidances and the China CDE computerized-systems guideline [4] [6] [12] [13] [14], and (4) draft guidance, including FDA-2024-D-4689 [1].
- Quality agreements. Apply the notice provision in Section 4.1 and use the single vendor request list in Section 8.3.
- Sidecar Audit Trail Architecture. Capture each raw inference and subsequent disposition under Section 5.2. Test reconstruction for human-reviewed, authorized automated and uncommitted outcomes. Record reviewer fields only when a person actually reviews the suggestion [12].
- Source lineage. Implement the investigator-control and working-copy procedure in Section 5.1.
7.3 Days 61 to 90: Golden-Set UAT, Adversarial Verification, and RBQM Metadata Monitoring
- Golden Benchmark Test Curation: Compile study-specific, anonymized golden benchmark test sets representing the full operational spectrum of trial data, including multi-site scans, degraded photos, and regional formatting variations.
- Adversarial UAT Execution: Execute the four targeted adversarial UAT checks across applicable tiers: (1) runtime provenance and hash verification (Tiers 2, 3, 4); (2) golden perturbation and locale testing (Tiers 3, 4); (3) refusal and uncertainty threshold testing (Tiers 1 Participant-Facing, 3, 4); and (4) uncommitted-draft isolation testing (Tiers 2, 3).
- Operationalizing RBQM Metadata Surveillance: Configure Risk-Based Quality Management monitoring dashboards to aggregate transaction metadata, tracking
AI Acceptance Rate,Human Override RateandMean Review Dwell Timeacross trial sites, subject to the reporting boundary in Section 5.3, as operational prompts to investigate potential automation complacency and regional distribution drift [2][21]. - Audit-trail export drills. Reconstruct an AI-assisted entry using the export procedure in Section 5.3.
8. Qualifying a Vendor Under a Confidentiality Agreement: An EClinCloud Case Example
When qualifying an enterprise eClinical platform vendor under a formal confidentiality agreement, sponsor QA auditors should evaluate whether the vendor's technical architecture supports regulatory compliance across discrete functional modules. Public documentation from EClinCloud (accessed November 13, 2026) illustrates how an integrated multi-module clinical software platform approaches task-scoped AI capabilities [35].
8.1 Public Trust Framework and Task-Scoped Architecture
On its public trust and compliance center, EClinCloud documents enterprise quality and security certifications, including ISO/IEC 42001:2023 (Artificial Intelligence Management System), ISO 27001 (Information Security Management), ISO 27701 (Privacy Information Management), ISO 9001 (Quality Management), ISO 20000 (IT Service Management), Cloud Security Alliance (CSA) STAR, and DJCP (China MLPS) [36]. In its public architectural statements, the company articulates a core governance principle:
"AI on the ECC platform is always tied to specific tasks — protocol parsing, OCR entry, document classification, risk scanning and Q&A — with humans reviewing outputs." [36]
According to company platform specifications, EClinCloud offers private / on-premises deployment options, with specific hosting arrangements confirmed with client teams during technical evaluation [35].
8.2 Module-by-Module Qualification Boundaries
This section applies the framework using EClinCloud's public statements. The verification items specify the evidence a sponsor would expect to review, and EClinCloud to provide, under a confidentiality agreement. The performance figures are company-reported. One eCOA figure is attributed to a pediatric Phase II study. The cited pages supply no denominators or evaluation methods for any of the figures.
Electronic Data Capture (EDC):
What EClinCloud States: EClinCloud states that large language model parsing of clinical study protocols can reduce initial study build effort by 60%, with generated visits, forms, and edit checks reviewed and confirmed by data managers before go-live [35][37]. The vendor also states that medical OCR operates across 1,000+ form types with 95%+ recognition accuracy on the platform page [35]. Its EDC page describes retrieval-augmented parsing [37]. The EDC page reports 95% recognition accuracy, parsing source records and syncing them straight into EDC, reducing per-form entry time from 8–10 minutes down to under 3 minutes and cutting source data verification (SDV) effort by 20%, incorporating automated PII redaction [37].
EClinCloud's Classification & Recommended Sponsor Verification: We would classify protocol parsing as Tier 2 and OCR source extraction as Tier 3 where authenticated confirmation precedes commit. During qualification audits, sponsors should verify: (1) whether the production EDC requires attributable human confirmation before committing OCR-extracted values and how it presents suggestions against the auto-fill guidance in Section 5.1; (2) that original unredacted source scans remain under investigator control; and (3) that raw OCR suggestions and human dispositions are recorded in an exportable sidecar audit trail.
Randomization and Trial Supply Management (RTSM):
What EClinCloud States: EClinCloud states that AI protocol parsing configures RTSM in one step, reducing setup effort by 10% to 20%, alongside intelligent supply alerts [38].
EClinCloud's Classification & Recommended Sponsor Verification: We would classify RTSM automated setup as Tier 2 pre-production configuration. During qualification audits, sponsors should verify: (1) that AI-generated randomization rules, kit types, and dispensing schedules are treated as uncommitted drafts requiring documented sign-off by a qualified biostatistician and supply lead prior to production release; and (2) that the system strictly prevents automated AI writes to live randomization tables or kit assignment algorithms during the active trial.
Electronic Clinical Outcome Assessment (eCOA):
What EClinCloud States: EClinCloud states that its eCOA module supports five assessment modalities: Electronic Patient-Reported Outcomes (ePRO), Electronic Clinician-Reported Outcomes (eClinRO), Electronic Observer-Reported Outcomes (eObsRO), Electronic Performance Outcomes (ePerfO), and Electronic Diaries (eDiary) [39]. The company reports that real-time camera image quality checks improved photographic pass rates in a pediatric Phase II study from 40% to approximately 95% while shortening rework cycles from 24–48 hours down to under 5 seconds [39]. It also reports automated drug-label text extraction with 95% accuracy in under 5 seconds and automated PII masking [39]. The platform includes a role-aware conversational assistant (LinBot) answering system and workflow questions [39].
EClinCloud's classification and verification. Advisory image quality checks are Tier 3 under Branch 2C. Label extraction is Tier 3 where the participant or site confirms the extracted value before it is saved. Verify attributable confirmation before commit, logging of raw extraction and confirmation, and the placement of suggested values against the auto-fill guidance in Section 5.1. LinBot’s workflow assistance, as described on the site, is Tier 1. If participants can use it, verify the navigation safeguards in Section 3.2, including applicable provider disclosures, refusal, escalation and site safety review [7].
Scientific & Clinical Trial Services (Human Specialist Boundary):
What EClinCloud States: EClinCloud's public services documentation lists Scale Management, Linguistic Validation (including translation, back-translation, and cognitive debriefing), and Rater Training as managed clinical trial services [40].
EClinCloud's Classification & Recommended Sponsor Verification: We note that these offerings represent specialized human clinical and linguistic services rather than automated software features. Sponsors should verify during service qualification that scale licensing permissions have been formally obtained from copyright holders and that linguistic validation workflows incorporate qualified target-language linguists.
Electronic Trial Master File (eTMF):
What EClinCloud States: EClinCloud states that its eTMF module provides automated document classification against the TMF Index alongside upload quality checks (detecting duplicate files, blank pages, blurred scans, and watermarks) and LinBot assistance, noting explicitly that "filing is confirmed by a person" [41].
EClinCloud's Classification & Recommended Sponsor Verification: We would classify this workflow as Tier 2 draft classification. Sponsors should verify: (1) that automated document index suggestions cannot transition an artifact to Filed status without authenticated human document specialist confirmation; and (2) that suggested classification tags and human confirmation actions are logged in the audit trail.
Clinical Trial Management System (CTMS):
What EClinCloud States: EClinCloud states that CTMS provides one-click AI project setup, RBQM-driven oversight and AI proofreading of monitoring reports. The platform page describes AI risk scanning [35] [42].
EClinCloud's Classification & Recommended Sponsor Verification: We would classify setup drafting as Tier 2, monitoring report proofreading as Tier 1, and AI risk scanning (platform page) as Tier 3. Sponsors should verify: (1) that risk signals and monitoring report drafts require authenticated human CRA and study manager sign-off before filing; and (2) that any AI risk model follows pre-defined recalibration boundaries under Pattern 2 change control.
Independent Review Committee (IRC), Resource Management (RM), and eLearning:
What EClinCloud states. EClinCloud lists IRC, Resource Management and eLearning as platform modules [35].
EClinCloud's Classification & Recommended Sponsor Verification: Sponsors should verify whether the IRC module incorporates any automated AI lesion scoring or operates purely as a blinded workflow management platform for human clinical reviewers, and confirm that eLearning scoring functions operate deterministically.
8.3 Recommended Audit Inquiries Before Study Go-Live
When conducting a formal qualification audit of EClinCloud or an equivalent eClinical software provider under a confidentiality agreement, sponsor QA auditors should request the following verification artifacts:
- ISO/IEC 42001:2023 Certification Package: Inspect the independent third-party audit report and Statement of Applicability for the vendor's Artificial Intelligence Management System, evaluating internal risk management procedures for algorithmic drift, data provenance, and human-in-the-loop controls [36].
- Contractual Model Governance Clauses: Confirm that the vendor Master Services Agreement or Quality Agreement includes: (a) the notice provision in Section 4.1; (b) regulatory inspection access to model architecture and training documentation where appropriate under EMA Reflection Paper Section 2.3.3.1 [9]; and (c) technical fallback rollback capabilities where feasible on self-hosted or supported vendor versions.
- Source lineage architecture. Verify the investigator-control procedure in Section 5.1 with technical data-flow diagrams for redacted clinical lab records or eCOA photos.
- Audit-trail export demonstration. Request a live reconstruction of an AI-assisted eCRF entry following Section 5.3.
- Hosting and Sovereign Boundary Verification: For multi-regional clinical trials subject to cross-border data transfer restrictions, inspect the vendor's private cloud or dedicated VPC configurations to confirm that clinical trial data and inference pipelines remain within designated national jurisdictions [35].
Sponsors seeking to discuss module-level qualification boundaries, audit trail schemas, and sovereign hosting architectures can contact EClinCloud through its product demonstration request page or general contact portal.
9. Methodology and Limitations
To ensure complete transparency regarding the evidence and analytical frameworks presented in this report, the following methodological boundaries and limitations are explicitly documented:
Regulatory text. The analysis uses the primary regulations, adopted guidance and drafts identified inline. China and EU implementation dates are those given in the cited material. Match the analysis to each study’s authorization, market and intended use.
Editorial convention. Binding rules are identified by their regulation or article numbers. Article 47 of the EU Clinical Trials Regulation tells sponsors and investigators to take appropriate account of the ICH GCP guidelines. The qualification tiers, change-control patterns, audit fields, numeric examples and schedule are EClinCloud recommendations [18].
Literature and primary evidence. Section 2 uses the published Shyr, Carroll, Yang, McKown, Karako and Gao, and Thierry reports. Sections 3 and 4 use the MHRA sandbox report, Yan’s QTL study and the regulatory sources cited there [10] [24] [25] [26] [27] [28] [21] [29].
Vendor Performance Metrics: All performance metrics and operational efficiency figures associated with EClinCloud software modules (such as OCR accuracy percentages, build time reductions, and photographic pass rates) are company-reported metrics from public documentation accessed on November 13, 2026 [35][36][37][38][39][40][41][42]. These metrics have not been independently tested or validated by regulatory authorities. Sponsors must independently validate software performance against their own study protocols, user requirements, and golden benchmark test suites.
Key Takeaways for Clinical Operations and QA Teams
- Identify binding rules and applicable guidance. Document the regulations, adopted guidelines and drafts relevant to the function. Article 47 of the EU Clinical Trials Regulation tells sponsors and investigators to take appropriate account of ICH GCP guidelines [18].
- Classify each workflow. Use the three-question screen in Section 3.1 to test unconfirmed writes, proximity to participants or decisions, and the influence and consequence ratings.
- Implement Deployment-Appropriate Version Control and Drift Guardrails: Require cryptographic SHA-256 digests of weights and container images for self-hosted open-weight models. For commercial cloud APIs, require pinned dated model identifiers, contractual advance notice for model deprecation, and periodic golden-set re-runs with pre-defined tolerance bands. For pivotal trials, EMA expects models and preprocessing pipelines used for hypothesis testing to be frozen and documented before database lock [9].
- Preserve the evidence for inspection. Use Sections 5.1 and 5.3 for investigator control, exports and the limit on individual dwell-time reporting.
Sources
1. U.S. FDA — Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products (Draft Guidance, Docket FDA-2024-D-4689, January 6, 2025; Draft — Not for Implementation)
2. International Council for Harmonisation (ICH) — Guideline for Good Clinical Practice E6(R3), Step 4 Final Consolidated Guideline (Adopted June 16, 2026; Principles & Annex 1 Finalized January 6, 2025; Annex 2 Adopted June 3, 2026)
3. European Medicines Agency (EMA) — ICH Guideline E6(R3) on Good Clinical Practice (Principles & Annex 1 Effective July 23, 2025; Annex 2 & Consolidated Version Effective January 15, 2027)
4. U.S. FDA — E6(R3) Good Clinical Practice (GCP) (Final Level 1 Guidance, Docket FDA-2023-D-1955, September 2025)
5. China National Medical Products Administration (NMPA) — Good Clinical Practice for Drugs (2026 Revision, Announcement 2026 No. 50, Jointly Issued by NMPA, NHC, NATCM, and NDCPA May 21, 2026; Effective September 1, 2026)
6. China Center for Drug Evaluation (CDE) — Notice 2026 No. 44 issuing the Guideline on Computerized Systems and Electronic Data in Drug Clinical Trials, effective September 22, 2026 (notice reprinted by China Food and Drug Net with the guideline PDF attached)
7. European Union, Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Official Journal L, July 12, 2024
8. European Union, Regulation (EU) 2026/1744 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 (Digital Omnibus on AI), Official Journal L, July 24, 2026
9. European Medicines Agency (EMA) — Reflection Paper on the Use of Artificial Intelligence (AI) in the Medicinal Product Lifecycle (EMA/CHMP/CVMP/83833/2023, Adopted September 2024)
10. Shyr C, Taylor R, Jagtap V, Cheng AC, Harris PA — Generative artificial intelligence implementation in REDCap, JAMIA Open, June 2026 (PMID 42369825 / PMC13302799)
11. U.S. FDA & European Medicines Agency (EMA) — Guiding Principles of Good AI Practice in Drug Development (10 Joint Non-Binding Principles, January 2026)
12. U.S. FDA — Electronic Systems, Electronic Records, and Electronic Signatures in Clinical Investigations: Questions and Answers (Rev. 1 Final Non-Binding Guidance, October 2024)
13. U.S. FDA — Digital Health Technologies for Remote Data Acquisition in Clinical Investigations (Final Non-Binding Guidance, December 2023)
14. U.S. FDA — Computer Software Assurance for Production and Quality Management System Software (Final Non-Binding Guidance, Docket FDA-2022-D-0795, Reissued February 3, 2026)
15. U.S. FDA — Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions (Final Non-Binding Guidance, Docket FDA-2022-D-2628, Reissued August 18, 2025)
16. U.S. FDA — Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations (Draft Guidance, Docket FDA-2024-D-4488, January 7, 2025; Draft — Not for Implementation)
17. FDA. E6(R3) Good Clinical Practice: Annex 2, Draft Level 1 Guidance. Accessed November 13, 2026.
18. European Parliament and Council — Regulation (EU) No 536/2014 on Clinical Trials on Medicinal Products for Human Use, and Repealing Directive 2001/20/EC, Official Journal L 158, May 27, 2014
19. European Medicines Agency (EMA) — Guideline on Computerised Systems and Electronic Data in Clinical Trials (EMA/INS/GCP/112288/2023, Published March 2023, effective six months after publication)
20. European Commission (DG SANTE) — Stakeholders' Consultation on EudraLex Volume 4 GMP Guidelines: Chapter 4, Revised Annex 11 and New Annex 22 on AI (Draft Consultation Closed October 2025; Draft Status Late 2026)
21. UK MHRA — AI Airlock Sandbox Phase 2 Programme Report (Supported by DHSC and RIO; 51 Applications, 7 Pilot Candidates; Published June 2026, Updated July 2026)
22. UK MHRA — AI Airlock Phase 3 Regulatory Sandbox Program Overview (Published October 6, 2026)
23. International Society for Pharmaceutical Engineering (ISPE) — GAMP Guide: Artificial Intelligence (July 2025) & GAMP 5 Second Edition Framework
24. Carroll A, Brown D, Silvio D, Maszatics J, Hanauer DA — Evaluating large language models for automated REDCap support ticket triage and response, JAMIA Open, June 2026 (PMID 42261354 / PMC13242945)
25. Yang Z — Machine learning for site risk prediction in clinical trials: development, external validation, and operational application in site qualification, International Journal of Medical Informatics, May 2026 (PMID 41741318)
26. McKown S, Arnold B, Boucher F et al. (ISOQOL TCA-SIG AI Working Group) — Use of AI within COA linguistic validation and eCOA migration processes: analysis and good practice recommendations, Journal of Patient-Reported Outcomes, February 2026 (PMID 41652288)
27. Karako K, Gao J — Artificial intelligence (AI)-assisted ultrasound in clinical trials: Endpoint automation, decentralized monitoring, and regulatory readiness, Drug Discoveries & Therapeutics, May 2026 (PMID 42002939)
28. Thierry A, Manson A, Podhaisky HP — How Digital Health Technologies Have Been Used to Support Regulatory Drug Approval, Therapeutic Innovation & Regulatory Science, Published Online July 2026; November 2026 Issue (PMID 42440213)
29. Yan L, Yu Z, Wu L, Liu R, Lin J — Optimizing Quality Tolerance Limits Monitoring in Clinical Trials Through Machine Learning Methods, Therapeutic Innovation & Regulatory Science, May 2025 (PMID 40000584)
30. European Union, Regulation (EU) 2017/745 of the European Parliament and of the Council on medical devices (MDR), Articles 62–82 and Annex XV: Clinical Investigations, Official Journal L 117, May 5, 2017
31. 21 CFR § 11.10 — Controls for closed systems (eCFR)
32. European Commission — Supporting the implementation of the AI Act with clear guidelines
33. European Union — Regulation (EU) 2017/746 on in vitro diagnostic medical devices, including performance studies
34. Medical Device Coordination Group (MDCG) & AI Board — MDCG 2025-6 / AIB 2025-1: Guidance on the Interplay Between the EU Medical Device Regulations and the EU Artificial Intelligence Act (June 2025)
35. EClinCloud — Unified Clinical Evidence Platform Architecture and Hosting Options (Accessed November 13, 2026)
36. EClinCloud — Trust & Compliance Framework (ISO/IEC 42001:2023 Certification, CSA STAR, DJCP China MLPS; Accessed November 13, 2026)
37. EClinCloud — Electronic Data Capture (EDC) Solution Specification (LLM Protocol Parsing & Medical OCR; Accessed November 13, 2026)
38. EClinCloud — Randomization and Trial Supply Management (RTSM) Solution Specification (Accessed November 13, 2026)
39. EClinCloud — Electronic Clinical Outcome Assessment (eCOA) Solution Specification (Five Modalities, Image QA & LinBot; Accessed November 13, 2026)
40. EClinCloud — Scientific & Clinical Trial Services (Scale Management, Linguistic Validation & Rater Training; Accessed November 13, 2026)
41. EClinCloud — Electronic Trial Master File (eTMF) Solution Specification (TMF Index Auto-Classification; Accessed November 13, 2026)
42. EClinCloud — CTMS: RBQM-driven oversight and AI report proofreading (Accessed November 13, 2026)