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

The paper ETHOS: Towards a Modular Ethics Framework for Clinical Multi-Agent Systems addresses the critical challenge of ensuring ethical compliance and safety when deploying complex, interacting Artificial Intelligence systems within clinical environments. As healthcare increasingly relies on interconnected multi-agent systems—where various AI components collaborate to support diagnosis, treatment planning, and resource allocation—the potential for unforeseen ethical failures increases exponentially. ETHOS proposes a novel, modular framework designed to systematically audit and govern the interactions between these agents, thereby establishing trust and accountability in high-stakes medical decision-making processes.

Core Architecture of the ETHOS Framework

ETHOS is fundamentally designed as a layered, modular architecture that allows for the integration of diverse ethical checks without requiring a complete overhaul when new agents or clinical protocols are introduced. The framework operates by intercepting the communication and decision pathways between collaborating AI agents, thereby providing an ethical choke point. This structure ensures that every proposed action or piece of derived information is vetted against established ethical principles before being presented to a human clinician. The system’s modularity allows researchers and developers to focus on specific ethical dimensions—such as bias detection or transparency logging—independently, facilitating iterative refinement. Key to its design is the concept of ethical provenance, which tracks not only the data source but also the ethical justification for every algorithmic step taken by any participating agent.

Ethical Dimensions and Principles

The framework systematically addresses several critical ethical dimensions that are often overlooked in purely technical AI deployments. These principles form the basis of ETHOS's decision-making logic, ensuring that clinical support remains patient-centric and equitable. The paper enumerates these core pillars:

1. Beneficence and Non-Maleficence: Ensuring that the collective output of the multi-agent system maximizes positive patient outcomes while rigorously minimizing potential harm. This involves continuous risk assessment across all proposed interventions.

2. Fairness and Equity: Actively detecting and mitigating algorithmic biases related to demographic variables (e.g., race, socioeconomic status, gender). ETHOS mandates that the system must demonstrate equitable performance across diverse patient cohorts before deployment.

3. Transparency and Explainability (XAI): Requiring that every decision or recommendation generated by any agent must be accompanied by a clear, auditable explanation of its reasoning. The framework emphasizes providing counterfactual explanations, detailing what inputs would need to change to alter the outcome.

4. Autonomy and Accountability: Defining clear lines of responsibility among the human clinician and the contributing AI agents. ETHOS ensures that the human remains in the loop, maintaining ultimate decision authority while documenting which specific agent contributed which piece of supporting evidence.

Operationalizing Ethical Auditing in Clinical Workflow

The practical implementation of ETHOS involves a multi-stage auditing process integrated directly into existing Electronic Health Record (EHR) systems. When a clinical consultation is initiated, the framework first performs a pre-flight ethical assessment, analyzing the initial data inputs for completeness and potential bias. As multiple agents contribute—for instance, one agent performing imaging analysis, another predicting risk scores, and a third suggesting treatment pathways—ETHOS continuously monitors the interaction flow.

The system utilizes a specialized module called the Conflict Resolution Engine. If two or more collaborating agents propose conflicting actions (e.g., Agent A suggests high-dose therapy based on biomarker X, while Agent B suggests conservative management due to patient comorbidities), this engine does not merely flag the conflict; it forces a structured deliberation. This deliberation requires all agents to submit a detailed justification, which is then weighed against the established ethical principles of ETHOS, thereby guiding the human clinician toward the most ethically justifiable path forward. The framework thus transforms passive AI support into an active, ethically guided collaborative process.

1. Implementation of a Modular Governance Meta-Agent (ETHOS Architecture)

• Improvement: Decouple ethical oversight from the core logic of individual agents by implementing an external, modular governance layer that interfaces with the system's tool registry and response pathways.

• Capability: Allows for the rapid updating of safety protocols and ethical policies across a multi-agent system (MAS) without requiring retraining or architectural changes to the underlying specialized models.

2. Deterministic Tool-Boundary Validation (Pre-specified Checks)

• Improvement: Integrate automated, deterministic checks at the point of tool invocation to validate data integrity.

• Capability: The system can automatically detect and flag technical failures such as:

• Segmentation Truncation: Identifying if an organ mask touches the image boundary, rendering volumetric measurements unreliable.

• Artifact Detection: Identifying metallic implants or focal contrast accumulation that distort attenuation-based biomarkers.

• Temporal Discordance: Flagging mismatches between the timing of laboratory results (e.g., blood work) and imaging scans to prevent clinically invalid indices.

• Evidence Grounding: Verifying that retrieved biomedical literature substantively supports the specific claim being made, preventing illusory grounding.

3. Statistical Out-of-Distribution (OOD) Monitoring

• Improvement: Embed model-agnostic OOD detectors (using Mahalanobis distance or k-nearest-neighbor distance) into the predictive tool pipeline.

• Capability: The system can quantify the distance between a current patient case and the model's training distribution, triggering an automatic abstention when a case is too novel to ensure safe decision-making.

4. Two-Tiered Contextual Reasoning Oversight

• Improvement: Implement a dual-level adaptive review process consisting of sub-agent self-interrogation and cross-agent synthesis review.

• Capability:

• At the Sub-Agent level: Forces individual agents to pause and assess if a specific patient’s unique clinical context requires additional evidence or a revision of their initial rationale.

• At the Orchestrator level: Detects contradictions or omissions that only emerge during synthesis (e.g., when one agent's imaging findings conflict with another agent's laboratory findings).

5. Iterative Ethics-Critic Adjudication

• Improvement: Deploy a final-stage LLM-based Ethics Critic that evaluates the compiled response against a structured rubric (Beneficence, Non-maleficence, Guardrails, and Safety-Trust-Patient frameworks).

• Capability: The system can act as a programmatic safety gate that:

• Forces the MAS into a bounded iterative loop (e.g., up to 5 rounds) to revise responses that fail ethical or safety thresholds.

• Completely suppresses and blocks the delivery of a response to the end-user if it cannot satisfy predefined safety criteria, replacing it with an explanatory notice of unresolved concerns.

The rapid adoption of large language models has enabled the development of clinical multi-agent systems (MAS) capable of integrating multimodal patient data and supporting increasingly complex clinical decision-making. However, the deployment of these systems in real-world healthcare settings raises critical ethical concerns related to safety, fairness, accountability, transparency, and patient trust. While numerous organizations, including the World Health Organization, the National Academy of Medicine, and the FUTURE-AI consortium, have proposed ethical frameworks and governance principles for healthcare AI, these efforts remain largely conceptual. To address this challenge, we present ETHOS (Ethics and Trust through Hierarchical Oversight System), a modular ethics framework designed as a governance meta-agent that can be integrated with any existing multi-agent system without requiring changes to its underlying architecture. ETHOS translates stakeholder-informed ethical requirements into executable runtime oversight through a layered governance approach consisting of deterministic checks, contextual reviews, and a final ethics critic. These components continuously evaluate intermediate reasoning steps and final outputs, enabling the system to identify ethical risks, request revisions, or suppress responses that fail predefined safety and trustworthiness criteria. We demonstrate ETHOS within a hepatology clinical decision-support MAS. Results show that ETHOS improves decision reliability by detecting incomplete, inconsistent, or out-of-scope evidence and appropriately increasing abstention when safe recommendations cannot be supported. By embedding ethical governance directly into system operation, ETHOS provides a practical and auditable mechanism for transforming high-level AI ethics principles into deployable safeguards.