AI Governance, Privacy, and Data Security in elluminate

Note: The agreements between eClinical Solutions and each sponsor govern the use of study data. For questions about a specific agreement, contact an eClinical Solutions representative.

AI Governance Principles

The following principles apply to every AI capability in elluminate:

Purpose-built: Each AI capability is designed for a specific clinical workflow, with a defined intended purpose and bounded scope.
Assistive: AI supports expert work and does not replace human judgment or accountability. AI outputs require user review before use.
Proportionate: Risk controls are applied based on the impact and context of each capability.
Governed by design: Governance, validation, and oversight are design inputs for every capability rather than additions after release.
Transparent: Product documentation describes the intended use, data inputs, outputs, and limitations of each capability. Detailed model designs remain proprietary to eClinical Solutions.

Controls Across the AI Lifecycle

STAGE	CONTROLS
Development	Governed development under a validated software development life cycle (SDLC), source control and code reviews, independent training and test datasets, subject matter expert (SME) involvement, and validation aligned to intended use.
Use	Human-in-the-loop workflows, explainable outputs, confidence indicators, user decision points, and access controls based on user permissions.
Audit	Traceability of AI versions, configuration, outputs, and user actions, with ongoing monitoring for consistency and performance drift.

Data Privacy and Use

Clear data boundaries apply to all elluminate AI capabilities:

Sponsor study data remain customer-controlled and are used only to deliver contracted functionality.
Study data are not used to train or improve shared or generalized AI models.
Study data are not exposed to public or consumer AI services.
Outputs generated from sponsor inputs are owned by the sponsor and remain inspectable.
These boundaries are enforced through technical, contractual, and operational controls.

Use of Study Data for Model Training

Sponsor study data are not used to train AI models. Study data are used at runtime only, as context to generate outputs for the study in which the data reside. Runtime use of study data is separate from model training.

Generative AI capabilities: Capabilities such as Data Advisor and AI Risk Statement generation use pretrained foundation models. eClinical Solutions configures the instructions, context, and tools for these capabilities and does not fine-tune the underlying models with study data.
Purpose-built models: Models such as Mapping Automation and eIQ Data Classification were developed using de-identified historical clinical trial data. Use of de-identified data for model development is subject to client permissions and contractual agreements.
User feedback: Thumbs up and thumbs down ratings, such as those in Data Advisor, are used to evaluate model performance.

Data Processing Location

Study data, prompts, and AI outputs remain within eClinical Solutions AWS infrastructure and the elluminate platform environment. Capabilities that use foundation models access those models through Amazon Bedrock within this environment. For example, Data Advisor uses Anthropic Claude models. Purpose-built models developed by eClinical Solutions, such as Mapping Automation, run within eClinical Solutions infrastructure. Study data are not sent to external systems, and Data Advisor does not use web search or other external sources.

Data Retention by Capability

CAPABILITY	SESSION AND HISTORY	ACCESS AND AUDIT
Data Advisor	Context is maintained for the active visualization or listing during the current session and is cleared when the page is refreshed or the session times out or ends. Context is not retained across sessions.	Requests and responses are recorded for audit and observability. Logs apply PII redaction and hashed identifiers.
elluminate Assist	Up to 10 previous chats are stored for each user, and each chat retains prompts from the last 30 days.	Chat history and returned data are limited by user privileges and data blinding settings.
Study Advisor	Uploaded study documents and previous document versions are retained in Study Advisor. Accept and reject decisions for Review Objective suggestions are retained across protocol versions.	Rejected suggestions record a Rejection Reason.
Mapping Automation	Recommendations are stored in the specification Variable List.	The audit log tracks automatic domain and codelist creation, accepted recommendations, rejected recommendations, modified recommendations, and applied recommendations.
AI Risk Statement generation	Generated statements are stored with the risk in the RACT assessment.	All AI-generated actions are logged for auditing.
eIQ Review	Model results, review status, and feedback are stored with the study in Data Central.	Review status is tracked for each record, and model run status displays for each model.

Data Access and Security

Access Controls

AI capabilities operate within the study context and permissions of the signed-in user. The following controls apply:

Role-based access control (RBAC): AI capabilities retrieve only data the signed-in user is authorized to access. Data access credentials are scoped to each signed-in user.
Study-level isolation: Data access is restricted to the originating study and tenant. AI capabilities cannot access data across studies or tenants that the user is not authorized to view.
Data blinding: Study blinding rules apply to AI outputs, so masked data remain protected.
Controlled tool access: Generative AI capabilities access data only through defined application tools and APIs. In Data Advisor, user input is not passed directly into tool calls.

Request Protection and Guardrails

Requests to Amazon Bedrock use encrypted, access-controlled connections, and activity is logged. Guardrails apply at the application level and the model service level, including the following:

Restriction of each capability to its intended clinical use cases, with safe refusal of unsupported or unsafe requests.
Input filtering, output constraints, and validation of outputs for structure and completeness.
Content filtering, sensitive data handling, and prompt injection protection.
Protection of internal system instructions and logic.

Study Data Modification

AI capabilities do not modify study data:

elluminate Assist guardrails block queries that update, alter, or delete data.
Data Advisor does not create queries and does not create issues without user confirmation.
eIQ Review flags records for review and does not change the underlying data.
Study Advisor, Mapping Automation, and AI Risk Statement generation produce suggestions that users accept, edit, or reject.

Output Accuracy and Reliability

Grounding AI Outputs in Study Data

Generative AI capabilities in elluminate base each output on the study data, metadata, and documents available to the user. The following methods support accurate, verifiable outputs:

Grounding in study context: Retrieval-augmented generation (RAG) supplies study-specific data, metadata, schema information, Review Objectives, and documents before a response is generated.
Separation of retrieval and interpretation: Tools retrieve data and perform computations, and the language model interprets and summarizes the results. Data retrieval and computations are deterministic.
Validation of generated output: elluminate Assist validates generated SQL before execution. Data Advisor validates outputs for structure and completeness. AI Risk Statement generation enforces the 'if / due to / then' structure.
Bounded scope: Each capability is limited to supported data and intended tasks. When unfamiliar variables or missing descriptions create ambiguity, Data Advisor asks clarifying questions and states assumptions.
Traceable results: Outputs display the basis for results, such as the SQL query and plain-language explanation in elluminate Assist and the Protocol Reference and Rationale for Study Advisor suggestions. Data Advisor insights can be validated against the underlying study data within the platform.

eIQ Review models and eIQ Data Classification analyze study data and metadata directly and do not generate text. These models can flag records that do not require action or miss records that do. Confidence indicators and similarity scores help reviewers prioritize and assess results.

Output Errors and Limitations

AI outputs can contain incorrect interpretations, omissions, or unsupported statements, and output quality depends on the available data and context. Some eIQ Review models are tuned to minimize false alarms, and others are tuned to detect more anomalies. No model identifies every relevant record.

Important: Users remain responsible for checking important findings against the underlying study data and applying professional judgment before taking action.

Response Variability

Generative AI responses can vary in wording, emphasis, or investigation path, even when the underlying data are unchanged. In Data Advisor, core findings remain consistent when based on the same data, inputs, and user role. AI Risk Statement generation provides alternative suggestions by design. eIQ Review models produce consistent results when run on the same data.

Validation and Monitoring

Validation by Capability

Each AI capability is validated against its intended use before release. Validation approaches differ by capability:

CAPABILITY	VALIDATION APPROACH
Data Advisor	Evaluation on a structured test set of Data Central listings and visualizations, comparison across multiple foundation models, rubric-based scoring calibrated against SME review, and scenario-based QA and regression testing that includes access boundaries and security behaviors.
elluminate Assist	Testing against a set of gold-standard questions developed with SMEs, run on anonymized studies from diverse therapeutic areas. Generated results are compared with expert-annotated SQL results, explanations are manually validated, and guardrails are tested with data modification requests.
eIQ Review	Separate training, validation, and test datasets, with ground truth established by independent data review SMEs. Software quality assurance testing includes reproducibility, failure scenarios, performance, and comparison of model output with results displayed in elluminate.
Mapping Automation	Evaluation on a held-out set of studies not used for training, with all outputs subject to expert review in a human-in-the-loop workflow.
AI Risk Statement generation	Evaluation against SME-created gold-standard Risk Statements across low, medium, and high complexity levels for factual alignment, structure, and terminology.

Post-Release Monitoring

Production eIQ Review models are monitored for input data quality, data drift, and changes in the percentage of flagged records.
Data Advisor agent executions are tracked for observability and performance, and evaluations are reapplied so changes in output quality are identified and addressed.
The elluminate Assist validation question set is run again when underlying models or supported domains change.
User feedback, such as thumbs up and thumbs down ratings, informs ongoing evaluation.

Human Oversight

User Decision Points

Every AI capability includes a user decision point before AI output is used:

CAPABILITY	USER DECISION POINT
Study Advisor	Accept, edit, or reject each suggested Review Objective, and click Add to Configuration to apply suggested study configuration values.
Data Advisor	Review summaries, insights, and recommendations, and confirm all issues before creation.
elluminate Assist	Review the SQL explanation and returned data, and save results as a custom listing when appropriate.
eIQ Review	Review flagged records, mark records as reviewed, provide feedback on predictions, and create issues.
eIQ Data Classification	Review concept tags and similarity scores before relying on a match.
Mapping Automation	Accept, reject, or update each recommendation in the Variable List.
AI Risk Statement generation	Review and edit the generated Risk Statement. A disclaimer displays until the content is edited.

Enablement and Control

AI capabilities are not enabled by default. Each capability is enabled through URL settings, study settings, or user privileges, and some capabilities require an eClinical Solutions representative to enable them. Organizations control which studies and users have access to each capability. Available AI capabilities vary by elluminate release.

For a list of AI capabilities and their locations in elluminate, refer to the AI Capabilities in elluminate article.