Streamlining Clinical Trial Documentation Generation with RAG and LLMs

clario · Healthcare · 2025 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Clinical-trial documentation generation

Clario needed to streamline generation of Business Requirement Specification documents and related clinical-trial documentation derived from medical imaging charters, a weeks-long manual process prone to errors, inconsistencies, redundancy, and downstream study delays.

Constraints:

• Clinical-trial documentation is regulated, sensitive, and tied to central review of medical images.
• The BRS, training slides, and ancillary documents are largely derived from the Charter but require reformatting and rephrasing.
• The solution is described as a prototype being productionized in 2025, not a long-running production system.
• Data needed to stay within the AWS ecosystem for security and compliance confidence.

Architecture

Clario and AWS built a RAG prototype for BRS generation. Charter-derived documents are prepared by an on-premises script, chunked, embedded with Amazon Titan Text Embeddings, stored in Amazon OpenSearch Serverless, retrieved by a workflow engine, and passed to Claude 3.7 Sonnet on Amazon Bedrock. A business requirement writer reviews generated answers in the UI, and final documents are written back to Amazon S3.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  medical_imaging_charter["Medical imaging charter"] --> on_premises_preparation_script["On-premises preparation script"]
  on_premises_preparation_script --> amazon_titan_text_embeddings["Amazon Titan Text Embeddings"]
  on_premises_preparation_script --> amazon_s3_document_storage["Amazon S3 document storage"]
  amazon_titan_text_embeddings --> amazon_opensearch_serverless["Amazon OpenSearch Serverless"]
  amazon_opensearch_serverless --> workflow_engine["Workflow engine"]
  workflow_engine --> amazon_bedrock_claude_3_7_sonnet["Amazon Bedrock Claude 3.7 Sonnet"]
  workflow_engine --> brs_review_interface["BRS review interface"]
  brs_review_interface --> amazon_s3_document_storage

Components

Data flow

Charter-derived documents are prepared for upload, chunked, embedded, and stored in OpenSearch Serverless. The workflow engine retrieves relevant charter chunks, calls Claude 3.7 Sonnet through Amazon Bedrock to generate BRS answers, presents the output in a UI for writer review, and stores final documents in Amazon S3 for future workflows.

Why this architecture

The architecture uses RAG and serverless AWS services because the downstream BRS content must stay grounded in medical imaging charters, scale with demand, and remain inside a secure AWS environment while still leaving a business requirement writer in the loop for final document quality.

Models used: Claude 3.7 Sonnet · Amazon Titan Text Embeddings · Amazon Bedrock foundation models

Scale

• Clario has supported more than 26,000 clinical trials and over 700 regulatory approvals across more than 100 countries.
  “supported clinical trials more than 26,000 times with over 700 regulatory approvals across more than 100 countries.”article, stated
• The existing documentation process could take weeks.
  “the time-consuming process of generating documentation for clinical trials, which can take weeks.”article, stated
• Clario is working to productionize the prototype in Medical Imaging during 2025.
  “Clario is working to productionize the solution in their Medical Imaging business during 2025”article, stated

Alternatives considered (roads not taken)

Manual reformatting and rephrasing from the Charter

Rejected because: The manual process was time-consuming and risked errors, inconsistencies, redundancy, and trial delays.

Ungrounded generation

Rejected because: The workflow retrieves relevant charter content from OpenSearch before generating answers, keeping outputs tied to source documents.

Fully automated final documentation

Rejected because: The source keeps a business requirement writer in the review path before final documents are produced.

Patterns

rag-retrieval (grounding)

The BRS generation workflow retrieves semantically relevant charter chunks from OpenSearch before invoking Claude.

human-in-the-loop-gate (quality control)

A business requirement writer reviews generated answers to produce the final clinical-trial document.

tool-use-orchestration (workflow orchestration)

A workflow engine coordinates vector retrieval, Bedrock calls, UI output, and final S3 document writing.

privacy-preserving-processing (security)

The solution keeps Clario information within AWS infrastructure using Bedrock security features.

Tradeoffs

Use RAG over charter-derived documents.

Chose OpenSearch vector retrieval plus Claude generation over Pure LLM generation from prompts.

Rationale: The BRS content is largely charter-derived, so retrieval grounds generation in the trial-specific source document.

Use serverless AWS services.

Chose Amazon S3, OpenSearch Serverless, and Amazon Bedrock over Self-managed vector and model infrastructure.

Rationale: The source emphasizes serverless scalability, modular integration, and keeping sensitive data inside AWS.

Prototype first before production rollout.

Chose AWS-developed prototype with 2025 productionization over Immediate full production deployment.

Rationale: The case reports a prototype that demonstrated benefits and is being productionized in the Medical Imaging business.

Gotchas (what bit them)

Manual document derivation can create clinical-trial delays.

Root cause: BRS and related documents are derived from Charters but require extensive reformatting and rephrasing.

Resolution: Automate BRS drafting with a RAG workflow while retaining human review.

Sensitive clinical data needs infrastructure isolation.

Root cause: Clinical-trial documentation contains information requiring confidentiality, integrity, and compliance confidence.

Resolution: Keep data within AWS infrastructure and use Bedrock security features.

Evaluation

The AWS/Clario post reports prototype benefits qualitatively: improved accuracy, scalability/flexibility, and security, but it does not provide a quantitative benchmark harness.

Metrics: Documentation generation previously could take weeks. · Clario-supported clinical trials: more than 26,000. · Regulatory approvals supported: over 700. · Countries supported: more than 100.

Outcomes

• The prototype streamlined BRS generation.
• It reduced risk of translation errors and inconsistencies.
• It demonstrated serverless scalability/flexibility and secure AWS-contained processing.
• Clario is working to productionize it in Medical Imaging during 2025.

The case reports prototype benefits and 2025 productionization plans, not production traffic or latency metrics.