pooja-k-swamy/clinical-protocol-review

Purpose and Runtime

Clinical Protocol Draft Assistant is a Python 3.8+ Streamlit application that generates or ingests clinical-trial protocols, obtains LLM-based expert review, and produces amendment-risk findings plus a quality score. It requires an OpenAI API key and uses LangChain-based agent behavior. Primary UI/runtime entry point: streamlit_app.py.

User Workflow Orchestrator

The Streamlit app is the control plane: users enter study parameters to draft a protocol or upload an existing PDF, TXT, or Markdown document; once protocol text is available, they trigger review and receive consolidated feedback, risk assessment, recommendations, and score.

• streamlit_app.py

• utils/document_processor.py

Protocol Draft Generation

The generation stage turns user-provided study parameters and a guideline template into an initial protocol draft through an LLM. ICH-oriented template content is a key prompt/input boundary rather than application code.

• agents/protocol_generator.py

• templates/ich_templates/ich_template_v1.md

Protocol Content Access Boundary

The MCP-inspired server provides structured protocol access to review agents, intended to expose targeted sections such as objectives or inclusion criteria instead of requiring every agent to process unstructured full text. This is the repository’s abstraction boundary between protocol content and reviewers; the README notes that the MCP model is currently planned for deeper implementation.

• mcp_interface/protocol_server.py

• mcp_interface/review_tools.py

Multi-Agent Clinical Review

Three role-specific LLM agents independently assess the protocol. Their perspectives divide responsibility across scientific rigor/feasibility/safety (PI), site execution and patient management (site physician), and regulatory, ethics, and data integrity (health authority). Their outputs feed the downstream consolidation and risk stages.

• agents/pi_agent.py

• agents/site_physician_agent.py

• agents/health_authority_agent.py

Risk and Scoring Pipeline

Consolidated agent feedback is interpreted as potential amendment triggers. The risk assessor emits severity—Low, Medium, or High—along with rationale and recommended remediation. The scoring engine converts identified risk severity into an overall numerical protocol score, where higher indicates fewer or less severe issues.

• utils/risk_assessor.py

• utils/scoring_engine.py

External Dependencies and Boundaries

The system runs locally under Streamlit, with Python dependencies defined in requirements.txt. OpenAI is the material external LLM service, configured via OPENAI_API_KEY; LangChain mediates agent/LLM behavior. Uploaded documents cross an ingestion boundary at the document processor, while generated and uploaded protocol text converge before MCP-mediated review.

• requirements.txt

• utils/document_processor.py

• mcp_interface/protocol_server.py