AWS Architecture Blog·August 19, 2026Architecting AI Agents for Clinical Trial Eligibility with AWS Bedrock
This article outlines the architecture for an AI-powered system designed to automate and accelerate clinical trial eligibility and safety screening. It details how AWS services like HealthLake and Bedrock AgentCore can be combined to ingest fragmented patient data, apply complex protocol criteria using multi-step AI agents, and provide human-in-the-loop oversight for critical clinical decisions, ensuring scalability and compliance.
Read original on AWS Architecture BlogThe article presents a robust system architecture for an AI-powered Clinical Trial Eligibility and Safety Agent built on AWS. This system aims to address the significant challenges in clinical trial enrollment, primarily stemming from fragmented data sources and the manual, time-consuming process of patient record review. By leveraging AI agents and a human-in-the-loop approach, the architecture seeks to improve both efficiency and accuracy, while maintaining clinician control and auditability.
Core Architectural Components
The proposed architecture integrates several key AWS services to form an end-to-end pipeline. The foundation is laid by AWS HealthLake, which standardizes and stores clinical data in a FHIR-native format, unifying disparate records from EHRs, labs, and imaging. Amazon Bedrock AgentCore serves as the orchestration layer for multi-step reasoning, coordinating specialized AI agents. Amazon Bedrock Knowledge Bases store the trial protocols and criteria that the agents reference. The overall flow includes data ingestion, agent-based assessment, LLM-as-a-judge evaluation, and a clinician review dashboard with audit trails.
Specialized AI Agents and Workflow
The system employs three distinct agents, each handling a specific phase of the screening process within the Amazon Bedrock AgentCore Runtime:
- Pre-screening Agent: Handles initial eligibility checks (consent, age, geography, washout periods).
- Detailed Screening Agent: The core reasoning engine, evaluating inclusion/exclusion criteria against FHIR resources, identifying safety signals, and assessing comorbidity profiles. It provides structured determinations with evidence and confidence scores.
- Site & Enrollment Agent: Manages logistical aspects, matching patients to appropriate sites based on proximity, capabilities, and capacity.
Ensuring Safety and Compliance with Guardrails
All agents operate behind Amazon Bedrock Guardrails to enforce critical non-functional requirements such as PII/PHI filtering, content safety, grounding checks to prevent hallucinations, and strict adherence to defined topic boundaries, which is crucial in healthcare applications.
Human-in-the-Loop and Continuous Improvement
A crucial aspect of this architecture is the Human-in-the-Loop (HIL) model. AI-generated decisions undergo a scoring process via Amazon Bedrock AgentCore Evaluations, which use an LLM-as-a-judge approach to assess clinical accuracy, operational effectiveness, and safety compliance. High-confidence decisions proceed, while lower-scoring or flagged cases are routed to a tiered clinical review structure (PIs, study coordinators, medical directors). Clinicians retain full override capability, and their corrections are captured to expand the ground truth dataset, enabling continuous learning and refinement of the AI agents.