Source-linked AI summary
ETHOS: Towards a Modular Ethics Framework for Clinical Multi-Agent Systems
Rakesh Sharma, Sydney Pugh, Cameron Beeche, Pankhuri Singhal, Rachel Wu, Margaret Eby, Jeffrey Duda, James Gee, Kyra O'Brien, Hersh Sagreiya, Marina Serper, Victoria Gershuni, Angela Bradbury, Anurag Verma, Eric Eaton, Kevin B. Johnson, Walter Witschey
TL;DR
Clinical multi-agent systems need executable, application-specific ethics governance beyond existing conceptual frameworks. ETHOS adds modular runtime oversight and, in a hepatology system, improved diagnostic measures while increasing abstention when evidence was insufficient.
Problem
Existing healthcare AI ethics frameworks provide conceptual categories but lack executable, application-specific runtime enforcement for clinical multi-agent systems.
Method
ETHOS is a modular governance meta-agent that applies stakeholder-informed deterministic checks, contextual reviews, and final ethics adjudication across clinical multi-agent systems.
Results
Across 150 decisions, ETHOS improved sensitivity on all three ICD-10 tasks and decision accuracy on two, increasing indeterminate decisions from 26.7% to 39.3%.
Takeaways & Limitations
ETHOS selectively withholds answers when required evidence is missing or incorrect, while leaving abstention unchanged when multimodal evidence is complete and consistent.
Takeaways & Limitations
Conclusions are bounded by evaluation of one MAS on a small manually selected cohort, unreviewed ETHOS findings and adjudications, and a subset of stakeholder concerns.
Abstract
from arXiv · showhide
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.
1. Introduction
Clinical multi-agent systems integrate multimodal information through specialized agents, but existing healthcare AI ethics frameworks remain largely conceptual and do not directly enforce requirements during system operation. ETHOS addresses this gap as a modular, stakeholder-informed governance meta-agent that applies runtime checks across clinical MAS components and is evaluated in a hepatology decision-support system.
- Motivation: Clinical multi-agent systems decompose complex clinical queries into steps performed by autonomous agents using specialized tools to integrate diverse patient information.Examples of tools include image segmentation and risk prediction models.
- Motivation: Existing healthcare AI ethics frameworks provide conceptual guidance, but the introduction identifies a need to enforce ethical requirements directly while a MAS runs.Referenced frameworks include guidance from the World Health Organization, FUTURE-AI, the National Academy of Medicine, and the American Medical Association.
- Design requirements: Runtime ethics enforcement must assess intermediate and final outputs, attach concrete checks to relevant agents and tools, and connect findings to consequences.These requirements translate stakeholder-defined ethical concerns into executable system behavior during inference.
- ETHOS framework: ETHOS is a modular ethics meta-agent that reuses core checks across MAS agents and tools while application-specific configuration supplies their scope and parameters as executable governance policy.The checks are informed by ethical concerns raised by clinical stakeholders through iterative co-design.
- Evaluation: ETHOS is evaluated in a hepatology MAS combining EHR laboratory results, CT images, and clinical practice guidelines to support disease screening, severity and risk assessment, and treatment planning.The application provides guideline-informed diagnosis and treatment support using multimodal patient data.
2. Related Work
Prior clinical and agentic AI ethics work largely defines principles without specifying deployable runtime checks, while technical enforcement systems remain limited in scope. ETHOS is presented as addressing this gap by translating ethical guidance into runtime, multimodal enforcement.
- Declarative ethical guidance: Existing clinical AI guidance, including WHO, FUTURE-AI, and National Academy of Medicine efforts, remains largely declarative rather than operational.These frameworks provide ethical constructs, checklists, or codes of conduct but do not specify how deployed systems should be checked against them.
- Unresolved clinical risks: Justice, explicability, and human-AI interaction remain comparatively underexplored, with little consensus on implementing clinical AI ethics principles.A benchmark also found that direct LLM-generated clinical recommendations risked severe patient harm in up to 24.6% of cases, mostly from errors of omission.
- Agentic AI ethics: Conceptual work on agentic AI argues that broader agency and deeper human-technology entanglement amplify established concerns around transparency, fairness, non-maleficence, accountability, and privacy.This literature identifies which principles agentic systems strain but does not propose a mechanism for enforcing them at runtime.
- Runtime enforcement: GuardAgent and AgentSpec demonstrate that externally specified rules can be enforced at runtime for LLM agents, but their focus is narrower than stakeholder-elicited clinical ethics concerns.GuardAgent generates and executes guardrail code for inputs and outputs, while AgentSpec compiles natural-language rules into runtime constraints that can intervene before actions complete.
- Research gap: The literature contains substantially more ethical guidance than work translating that guidance into runtime, multimodal enforcement, motivating ETHOS as a framework designed to close this gap.ETHOS is positioned against the limited technical literature on enforcement, which is largely outside clinical multi-agent settings and framed around safety and access control.
3. Stakeholder Engagement
ETHOS was shaped through iterative mixed-methods engagement with ten healthcare stakeholders and by mapping their feedback to established healthcare AI ethics frameworks. The resulting requirements motivated modular oversight components tailored to fixed criteria, patient-specific reasoning, and complete clinical answers.
- Stakeholder Engagement: Ten experts across healthcare providers, operations, leadership, informatics, privacy, legal, and governance contributed stakeholder input through an iterative mixed-methods design.The engagement examined ethics’ impacts, benefits, and implementation challenges in diverse real-world clinical settings during MAS design and evaluation.
- Stakeholder Engagement: Stakeholder feedback was mapped to principles from FUTURE-AI, the National Academy of Medicine, the American Medical Association, and World Health Organization frameworks.Feedback was linked according to how frequently concerns arose across stakeholder meetings.
- Stakeholder Engagement: Requirements were not one-to-one with ETHOS components, motivating checks distributed across multiple system points.Table 1 maps stakeholder requirements to ethics principles, illustrative ETHOS review excerpts, and triggered ETHICS checks.
- Stakeholder Engagement: Fixed-criterion requirements enabled deterministic checks for temporal discordance between scans and laboratory values and for contrast-enhanced-study attenuation thresholds.These checks assess tool-input and tool-output quality before or independently of patient-specific reasoning.
- Stakeholder Engagement: Other requirements required patient-specific reasoning about whether reference-threshold populations were appropriate for the individual case.This example concerns demonstrating local clinical validity rather than verifying a criterion fixed in advance.
- Stakeholder Engagement: Requirements concerning AI communication could only be assessed after drafting a complete clinical answer.Communicating AI’s role in a clinical decision is treated as a property of the final response, not a single tool invocation.
4. Methods · 4.1. ETHOS
ETHOS is an external governance meta-agent that adds ethical and reliability oversight to clinical multi-agent systems without modifying their underlying architecture. Its layered process combines deterministic tool-boundary checks with adaptive contextual review and ethics adjudication before final outputs are delivered.
- 4.1. ETHOS: ETHOS provides external ethical and reliability oversight for clinical multi-agent systems without requiring changes to the underlying architecture.It operates as a governance meta-agent attached to an existing host system.
- 4.1. ETHOS: A lightweight connector links ETHOS to the host system’s tool registry, agent finalization stage, and response revision pathway.The planner, specialist agents, and tools remain unchanged.
- 4.1. ETHOS: ETHOS uses a layered governance architecture to organize its oversight functions.The supplied methods passages describe deterministic checks, contextual review, and ethics adjudication as complementary layers.
- 4.1. ETHOS: The pre-specified layer automatically performs deterministic safety and reliability checks at tool boundaries whenever associated tools are invoked.Each check produces a structured finding describing the detected issue.
- 4.1. ETHOS: Structured findings are appended to tool outputs and passed to the sub-agent LLM.This makes detected issues available within the sub-agent’s subsequent processing.
- 4.1. ETHOS: Before finalization, contextual review asks whether a clinical scenario requires scrutiny beyond the predefined rule set.When appropriate, the agent performs supplemental analyses or revises its rationale.
- 4.1. ETHOS: After agent outputs are aggregated, an additional review layer performs ethics adjudication.The supplied passage identifies this as a post-aggregation governance step, though its further evaluation details are truncated.
4.2. Pre-Specified Checks
ETHOS uses deterministic pre-specified checks at imaging-tool boundaries and across model applicability, evidence relevance, and temporal coherence. These checks identify unreliable inputs, out-of-distribution cases, illusory grounding, and mismatched laboratory-imaging timing.
- Imaging integrity: ETHOS applies deterministic imaging checks at tool boundaries to identify measurements or predictions derived from potentially unreliable segmentation masks.The checks target image and measurement integrity, including contrast-phase uncertainty and segmentation masks that contact the image boundary.
- Model applicability: An out-of-distribution check flags cases whose distance from a tool’s development cohort exceeds a threshold.The model-agnostic check operates on fixed-length representations of model inputs or outputs, including learned embeddings and hand-crafted features.
- Model applicability: ETHOS jointly evaluates CT foundation-model applicability using out-of-distribution, region-coverage, and region-specific classifier-head checks.The out-of-distribution component determines whether an input scan lies outside the model’s established training distribution.
- Evidence grounding: An evidence-relevance check classifies retrieved biomedical or guideline material as directly supportive, tangentially related, or non-supportive of its cited claim.The check is designed to mitigate illusory grounding when the host multi-agent system uses external evidence.
- Temporal coherence: A deterministic temporal-coherence check evaluates laboratory–imaging timing, including inputs to indices such as FIB-4 that combine imaging context with AST, ALT, and platelet values.ETHOS reports temporal offsets as informational findings.
4.3. Contextual Checks
ETHOS adds a contextual review layer because fixed deterministic checks cannot exhaustively cover all clinically relevant ethical and reliability considerations. Before a specialist sub-agent finalizes its output, ETHOS interrupts the process and prompts case-specific scrutiny of the patient’s clinical context.
- Contextual Checks: ETHOS uses contextual review to perform case-specific governance beyond a fixed library of deterministic checks.The layer addresses clinically relevant ethical and reliability considerations that pre-specified checks may not exhaustively cover.
- Contextual Checks: Before a specialist sub-agent finalizes its output, ETHOS returns it to the reasoning loop with a prompt to assess whether the patient’s context requires additional scrutiny.This directed interruption occurs immediately before finalization and extends review beyond pre-specified checks when warranted by the case.
4.4. Ethics Critic
The ethics critic is ETHOS’s final holistic review layer, adjudicating the compiled multi-agent response using the draft, accumulated governance findings, and retrieved clinical and ethics guidance. Its assessment supports a programmatically enforced release decision and can trigger bounded critique–revision when requirements are not met.
- Inputs and holistic review: The ethics critic reviews the compiled response using the host MAS draft, accumulated pre-specified and contextual findings, and retrieved case-relevant clinical and ethics guidelines.It is implemented as a language-model reviewer operating on these three inputs.
- Inputs and holistic review: The critic holistically evaluates whether the response satisfies ethical and safety requirements for clinical deployment, recording rationale, governance findings, guideline citations, and revision recommendations.This differs from preceding layers that address individual concerns.
- Release decision: Release approval requires both ethical pillars to meet predefined minimum criteria, no guardrail violations, and no unresolved safety-trust-patient concerns.The default threshold is 3 of 5 for each ethical pillar.
- Release decision: When approval is withheld, ETHOS returns the response and feedback to the host MAS for revision and re-evaluates the revised draft through a bounded critique–revision cycle.The critic’s output serves as the basis for this programmatically enforced release decision.
5. Hepatology Multi-Agent System
The hepatology MAS is a modular clinical decision-support system that orchestrates specialist agents to synthesize multimodal evidence for liver-disease assessment. ETHOS was evaluated on 50 held-out patient cases covering three hepatology-related diagnoses from single-time-point imaging data.
- System Architecture: The MAS supports outpatient and specialty hepatology assessment by decomposing clinician queries, routing subtasks to specialist agents, and synthesizing their findings.It operates where CT imaging and routine laboratory data are available.
- System Architecture: Two complementary imaging agents analyze CT data: RadX extracts segmented-organ and radiomic features, while Percival generates volumetric imaging-based risk estimates.RadX determines contrast phase, uses TotalSegmentator for abdominal-organ segmentation, and applies PyRadiomics; Percival uses a three-dimensional CT foundation model.
- Data and Evaluation: The development resource comprised 225,896 CT studies from 72,523 unique patients in the Penn Medicine Biobank.The data included longitudinal clinical records, laboratory measurements, imaging studies, and linked electronic health records, under Institutional Review Board protocol 813913.
- Data and Evaluation: 50 independent held-out patient cases were evaluated for hepatic fibrosis, liver cirrhosis, and fatty liver using CT, laboratory measurements, demographics, and imaging metadata from a single imaging time point.The diagnoses corresponded to ICD-10 codes K74.0, K74.6, and K76.0.
6. Results
ETHOS preserved the base MAS’s abstention behavior with complete, correctly matched inputs but increased abstention when required evidence was missing or incorrect. Across the full cohort, ETHOS increased abstentions while improving selected non-abstained performance metrics.
- Complete-input stratum: With complete inputs, ETHOS left abstention rates unchanged: 0% for K74.0 and K74.6, and 60% for K76.0.This stratum contained 15 decisions from 5 patients across 3 ICD codes.
- Inadequate-input stratum: ETHOS increased inadequate-input abstentions from 40% to 67% for K74.0, from 47% to 67% for K76.0, and from 33% to 53% for K74.6.Across the stratum, abstentions rose from 18 of 45 decisions (40%) to 28 of 45 (62%).
- Full-cohort performance: Across the full cohort, ETHOS increased overall abstentions from 40 to 59 decisions (26.7% to 39.3% of the cohort).The largest increase occurred on K76.0, from 10 to 21 abstentions.
7. Discussion · Supplementary Material
ETHOS provides modular, end-to-end runtime governance for a hepatology clinical multi-agent system, translating stakeholder-informed ethical requirements into operational oversight. Its benefits arise partly through targeted abstention on unsupported cases, while cross-application reuse remains limited by application-specific criteria and evaluation scope.
- 7. Discussion: ETHOS is a modular meta-agent framework for clinical multi-agent systems, informed by clinical AI governance frameworks and clinical stakeholder input.It was applied to a hepatology system combining EHR laboratory values, CT imaging, and retrieved clinical practice guidelines for liver-disease screening.
- 7. Discussion: ETHOS governed the hepatology system end to end, generating findings at tool-call boundaries, during cross-agent evidence compilation, and at later stages.
- 7. Discussion: The governed system performs better on answered cases partly because it abstains from hard or ambiguous cases that lack sufficient diagnostic evidence.The resulting accuracy gain is therefore partly a selection effect rather than evidence of improved reasoning.
- 7. Discussion: When required modalities were missing or incorrect, abstention rose from 40% to 62%, driven primarily by Imaging Appropriateness and Contextual checks.When modalities were present and correctly matched, ETHOS left abstention rates unchanged for every code.
- 7. Discussion: Attempting to govern a cognitive-impairment MAS exposed the need to separate cross-application ethical requirements from application-specific operational checks.Imaging-specific criteria placed in the shared ethics layer prevented any response from passing adjudication.
- 7. Discussion: The conclusions are limited by evaluation of one MAS on a small, manually selected cohort and by enforcement of only stakeholder concerns expressible as verifiable conditions.Reuse was demonstrated across the agents and tools of one system rather than across clinical applications, while diagnostic outputs were evaluated against ground-truth labels.