Permissions: All Data Central users can use elluminate Assist.

elluminate Assist is a generative AI chatbot that simplifies access to SDTM clinical trial data in Data Central by translating plain-language queries into SQL code and returning relevant results.

Access elluminate Assist

Important: Review the General Information, Liability Disclaimer, and Tips for Effective Use sections before using elluminate Assist.

Click the elluminate Assist icon in the master header.

The elluminate Assist window opens, displaying links to instructions, several pre-defined prompts, a link to view additional prompts, and an input field for entering a plain-language prompt.

Access Instructions

Click the Instructions button.

The Instructions page contains the following tabs:

About

Review the General Information and Liability Disclaimer sections before using elluminate Assist. These sections describe the functionality and limitations of using AI to access SDTM data.

Tips

Review best practices and usage guidance for: Access and user-specific information Querying and data specifications Ensuring accuracy Managing chats Historical prompts and responses Use the scrollbar to view all content.

Quick Prompts

Use pre-defined prompts to begin querying study data. Subject matter experts compiled a collection of commonly used study questions.

Click Close Instructions to close the Instructions page.

Ask Questions

Use pre-defined prompts or enter custom prompts to query SDTM data. Review the structure of the pre-defined prompts to help create effective custom prompts. After a prompt is submitted, elluminate Assist processes the query and generates results.

Prompt Guidelines

Follow these guidelines when entering custom prompts:

• Enter prompts as questions or direct instructions.
• Limit prompts to supported SDTM domains: AE, CM, DM, DS, EG, EX, LB, MH, PR, SV, and VS.
• Use precise wording to avoid ambiguous results.
• Break complex questions into smaller components. elluminate Assist retains previous prompts and responses during the session.
• For incomplete date variables, elluminate Assist assumes:
  • The first day of the month when the day is not specified.
  • January when the month is not specified.
• Prompts can contain up to 500 characters.

Important: elluminate Assist cannot modify or delete data. Do not submit prompts that request data modification or deletion.

Send a Custom Prompt

1. Enter the prompt in the input field.

2. Click the up arrow to send the prompt. Processing may take a few moments.

   

Use a Pre-defined Prompt

Use one of the displayed pre-defined prompts from the main window, or access additional prompts from the View More Prompts link.

1. From the main window, select one of the displayed prompts.

   • Alternatively, click the View More Prompts link to open the complete list of available prompts. The full list is also available from the Quick Prompts tab on the Instructions page.

2. Click the up arrow to send the prompt.

Interpret Results

After a prompt is submitted, elluminate Assist returns a structured response.

The response includes the following information:

• Breakdown of SQL Steps: Describes the SQL logic in plain language and explains how the returned data are generated.
• Open Result in a Listing: Opens the results in a draft custom listing that can be saved in Data Central. Refer to the Open in Custom Listing section for details.
• Sources: Displays the SQL query and SQL query results.

Tip: Review the SQL explanation and returned data to confirm accuracy before refining prompts or generating additional queries.

Open in Custom Listing

Open prompt results in a Draft Custom Listing, then save the listing in the Visualizations section of Data Central. Saved listings support the same functionality as other Custom Listings, including docking, editing, and filtering.

Open and Save Results as a Custom Listing

1. Click the Open Result in a Listing link.
2. Review the draft custom listing.
   • The draft listing supports viewing, exporting, and maximizing, but it cannot be docked.
     
3. Click the Save icon.
4. In the Save Custom Listing window:
   • Enter a name for the listing.
   • Select the destination folder.
   • Select the Public or Private radio button. The Public option is only available to users with Data Central Designer privileges.
   • If saving as Public, select the roles that can access the listing.
   • The scope defaults to the current study. To update the scope, select different Therapeutic Areas, Compounds, Programs, or Studies.
5. Click Save to confirm changes or Cancel to discard changes.

The saved listing appears in the Visualizations section and responds to Subject Filters and Data Cuts.

Edit Results

1. Open the custom listing from the Visualizations section.

2. Click the Edit icon.

   The Custom Listing Designer window opens and displays:

   • A preview of the listing
   • The SQL query in the SQL Editor

3. Modify the SQL query as needed.
4. Click Validate & Update Preview.
5. Review the updated preview and validation results.
6. Click Save to save the changes and exit Edit Mode.

Note: After the SQL query is modified, the Save button remains disabled until Validate & Update Preview is completed.

Chat Session Toolbar

The chat session toolbar provides access to chat management options.

• New Chat: Click to return to the main window and start a new chat session.
• Chat Histories: Click to view previous chat sessions and search chat history.
• Close: Click to close elluminate Assist.

View Chat History

elluminate Assist stores up to 10 previous chat sessions. Each session retains conversation history for 30 days.

1. Click the Chat Histories icon in the toolbar.
2. Select a previous session or search the chat history.

Note: Chat history and returned data are limited by user privileges and data blinding settings. Each session displays only data accessible to the current user account.

Prompt Best Practices

Use the following best practices to create clearer prompts and improve the accuracy of generated SQL and results.

• Break Complex Questions into Smaller Logical Components
  When a question involves multiple domains, timing logic, or derived calculations, separate the analytical components before creating the query. Instead of combining joins, filters, grouping, and calculations into a single prompt, isolate the core clinical condition first. This approach helps reduce overly complex SQL generation and lowers the risk of unintended joins or misinterpreted logic.

  Example:

  Step 1. Show me a list of records with Subject ID (from the DM domain), 
        Site, Adverse Event Term, Serious flag, and Treatment Emergent flag. 
        Do not group or count.
  Step 2. From this result, show me only the records with Serious Flag = Y 
        and Treatment Emergent Flag = Y
  Step 3. Count the distinct Subject IDs from this result and group it by Site

  Focus: Logical decomposition of a complex problem.

• Clearly Define Subject Population
  Clearly specify the intended subject population, such as Enrolled, Randomized, Treated, Completed, or Discontinued subjects. If the population is not defined, the chatbot may assume a broader dataset than intended. When calculating rates or percentages, explicitly define the denominator (the group being divided by) to ensure accurate counts, rates, and percentages.

  Example:

  • Show rate of Serious AEs among Enrolled subjects.
  • Calculate percentage of Treated subjects with at least one Treatment-Emergent AE.
  • Return count of Randomized subjects with ALT > 3x ULN.

• Explain Specialized Terminology When Needed
  When a question contains study-specific terminology, define the terms using measurable criteria. Do not assume the chatbot recognizes internal cohort names or derived logic. Instead, convert those terms into specific variables, operators, and thresholds to help generate accurate SQL logic.

  Example:

  • Show subjects in the 'High Risk Cohort' (defined as subjects where AESER = 'Y' 
          for at least one adverse event and LBSTRESN is greater than 
          (3 times the LBSTNRHI) for at least one LBTESTCD = 'ALT').
  • Identify 'Early Discontinuations' (defined as subjects where 
          DSDECOD = 'DISCONTINUED' and DSSTDY is less than 30).
  • Show subjects with 'Frequent Visit Changes' (defined as subjects with SV records 
          where the count of SVTESTCD = 'VISIT' is greater than 5).
  

• Specify Aggregation and Output Structure
  Clearly specify whether counts of records or distinct subjects are required, and define the expected output structure, such as one row per subject, one row per site, or grouped by country. Without this information, the chatbot may return results at an unintended level of detail. Clear grouping and aggregation instructions help improve result accuracy.

  Example:

  • Return one row per Site with count of distinct subjects with Serious AEs.
  • Show total AE records by Study and Country.
  • One row per Subject with maximum ALT value.

• Use Measurable Criteria Instead of Qualitative Terms
  Avoid vague phrases like 'high labs,' 'early in study,' or 'bad data.' These terms are subjective and cannot be reliably translated into SQL. Instead, use specific variable names, operators, and numerical thresholds, for example, LBSTRESN > 3 * LBSTNRHI or AESTDY <= 7. Well-defined, measurable criteria help improve SQL accuracy.

  Example:

  • Show subjects where LBTESTCD = 'ALT' and LBSTRESN > (3 * LBSTNRHI) and LBDY > 1.
  • List adverse events where AESER = 'Y' and AEREL = 'Y'.
  • Identify lab records where LBSTRESN >= (2 * LBSTNRHI) and LBSTNRHI is not null.

• Define Time Logic Explicitly
  When using timing concepts such as baseline, post-baseline, treatment-emergent events, or follow-up periods, clearly define the time window in the prompt. Specify whether the comparison is based on Study Day, First Dose Date, baseline value, or another date variable. Avoid vague phrases like 'after baseline' or 'during treatment' without defining the exact time range. Clear timing logic supports accurate record selection and reduces the risk of including pre-treatment data.

  Example:

  • Show adverse events that started on or after the subject’s first dose date.
  • Show lab results collected after Study Day 1 and compare them to each 
          subject’s baseline lab value.
  • List adverse events that occurred within 30 days after a subject’s last dose date.
  

  Note: Many SDTM datasets include a Baseline Flag (for example, LBBLFL = 'Y') that directly identifies baseline records and is the most reliable reference. If no Baseline Flag exists, baseline must be defined using study timing logic, such as the earliest record, Study Day 1 (or Day 0), or a visit labeled 'Baseline.' Always describe baseline using these study terms rather than assuming a specific visit number if no Baseline Flag exists.

• Review and Validate Output Carefully Before Refining
  Before entering follow-up prompts, review the output and SQL to confirm they meet expectations. Check subject counts, filters, domain joins, and derived logic. If something is incorrect, refine the prompt before adding complexity. Building on incorrect results can create compound errors.

  Example:

  • Explain the SQL logic used for this result.
  • Confirm whether distinct subjects are being counted.
  • What joins are used between AE and DM?

• Iterate and Refine Based on Output Review
  After generating an initial result, review both the output and SQL logic before expanding the query. Use follow-up prompts to adjust grouping, refine filters, add denominators, or clarify time logic. Iterative refinement improves accuracy and helps prevent added complexity based on incorrect interim results.

  Example:

  Step 1. List all Serious AEs with Subject ID and Site from the DM table.
  Step 2. Using the results above, filter the records to only display the 
        ones with AEs that are Treatment Emergent.
  Step 3. Count the number of distinct Subjects from this result by each Site ID.
  Step 4. Append a new column to this result to show the percentage of 
        the total number of enrolled Subjects per Site ID. 

  Focus: Progressive enhancement after validation.