AWS-based Trusted Research Environment 'NDS AI·RE' Secures Medical Data Security and AI Analysis Speed Simultaneously with Internet-Blocked VPC

Source Isolation via VPC Without Internet Gateway

NDS AI·RE is a Trusted Research Environment (TRE) solution for handling sensitive medical and research data, built on AWS infrastructure. The core concept is configuring a VPC without an Internet Gateway to fundamentally block all outbound internet paths. Isolation is guaranteed by the network structure itself rather than rule-based blocking, allowing only two external communication channels—a data storage path and a status publishing path—both of which terminate within the AWS internal network.

AI Controls Combined with the Five Safes Framework

This solution follows six design principles that add AI-Enabled to the Five Safes framework adopted by the UK Office for National Statistics (ONS) and others.

• Safe People: Access is restricted to authorized researchers via MFA authentication and Role-Based Access Control (RBAC)
• Safe Projects: Automatic deployment of independent analysis spaces per project and resource limit settings
• Safe Settings: Use of only verified tools within an isolated, internet-blocked environment
• Safe Data: Read-only mounts without original data duplication (Zero-Copy architecture)
• Safe Outputs: Result export only after sensitive information inspection and administrator approval
• AI-Enabled: AI agents operate only within the control scope of the above five safeguards

Controlled AI Agents and Automated Onboarding

AI Core is designed so that all model calls pass through a single LLM Gateway, enhancing security and audit trail management. Researchers input analysis objectives in natural language, and AI agents assist with data structure interpretation and code generation, with the generated code executing within the Enclave. Amazon Bedrock is used for model calls, eliminating the need to operate separate model servers.

Infrastructure management is performed using IaC (Infrastructure as Code) methods leveraging Terraform and Packer. When onboarding new institutions, the platform automatically executes Terraform, with rollback processing in case of failure to prevent incomplete states. This resolves the trade-off between security compliance and research speed (AI Transformation).