ETHOS: Towards a Modular Ethics Framework for Clinical Multi-Agent Systems

ETHOS translates abstract ethical principles into concrete, enforceable rules: Rather than remaining theoretical, the framework converts stakeholder-informed ethical requirements into executable checks that run continuously during system operation, addressing a gap where existing WHO and National Academy of Medicine guidelines have remained largely conceptual.

The framework uses a layered oversight approach to catch ethical problems at multiple stages: ETHOS employs three levels of evaluation—deterministic checks, contextual reviews, and an ethics critic—that examine both the intermediate reasoning steps and final outputs of clinical AI systems, enabling the system to request revisions or suppress unsafe responses before they reach patients.

ETHOS can be added to existing clinical AI systems without major redesign: As a "governance meta-agent," the framework integrates with multi-agent systems without requiring changes to their underlying architecture, making it potentially practical for hospitals and healthcare organizations to adopt without completely rebuilding their AI infrastructure.

The framework addresses multiple critical ethical concerns simultaneously: ETHOS is designed to evaluate safety, fairness, accountability, transparency, and patient trust—recognizing that responsible clinical AI requires oversight across multiple ethical dimensions, not just accuracy or efficiency.

Real-world testing demonstrates feasibility in hepatology decision-support: The authors demonstrated ETHOS within an actual clinical use case (hepatology), providing evidence that the framework can function in realistic healthcare contexts rather than remaining a theoretical proposal.