Clinical Harness for Governable Medical AI Skill Ecosystems

Technical Summary: Clinical Harness for Governable Medical AI Skill Ecosystems

Problem Statement
The paper identifies a critical translational gap in medical AI: while deep learning and large language models have advanced rapidly in specific tasks (e.g., imaging interpretation, risk prediction), their clinical deployment remains uneven and lacks evidence of improved patient outcomes. The core issue is organizational rather than purely technical. Current medical AI is typically built as isolated, task-specific models optimized for single endpoints. In contrast, clinical care is a longitudinal process involving multiple disciplines, visits, and interventions where evidence, uncertainty, and responsibility must persist over time. A single model cannot determine its own invocation timing, validate its inputs, manage handoffs to other models, or handle failure scenarios. Consequently, medical AI deployment is fundamentally a problem of capability organization.

Methodology and Framework
The authors propose a two-part framework to bridge this gap: Clinical AI Skills and the Clinical Harness.

1. Clinical AI Skills:
   The authors redefine the unit of AI capability from a "model" to a "skill." A Clinical AI Skill is a software-defined capability that performs a specific task within a defined clinical context. Unlike a raw model, a skill carries explicit metadata regarding:

   • Evidence and Boundaries: Intended-use populations, contraindications, and links to guidelines or validation studies.
   • Safety Constraints: Logic for withholding, falling back, or triggering human review when inputs are invalid or confidence is low.
   • Governance: Version control, audit trails, and revalidation requirements.
   • Standardization: Skills are encapsulated in a "skill card" (a technical contract and clinical instruction label) that documents interfaces, operating boundaries, evidence levels, and known failure modes.

2. Clinical Harness (Runtime Governance Architecture):
   The Clinical Harness is the infrastructure that registers, orchestrates, guards, and monitors these skills. It operates through four distinct layers:

   • Multilayer Safety Guardrails: Enforces input validation, hard clinical constraints, confidence monitoring, and blocks actions outside registered scopes. High-risk decisions (e.g., surgical planning) require clinician confirmation.
   • Skill Registry and Boundary Management: Maintains the "skill cards" to prevent unregistered or off-label execution, providing metadata for safety checks.
   • Orchestration and State Management: Uses deterministic clinical-pathway directed acyclic graphs (DAGs) and finite-state machines to schedule skill invocation. This replaces unconstrained natural-language planning with predefined pathways, ensuring intermediate decisions are reproducible and auditable via structured artifact contracts.
   • Audit and Continuous Monitoring: Records immutable logs of user identity, data hashes, outputs, guardrail events, and clinician actions to support traceability and trigger alerts for drift or safety events.

Exemplar Application: Osteoporosis Lifecycle
To demonstrate the framework, the authors outline an osteoporosis-oriented ecosystem comprising nine distinct skills (S1–S9) that span the entire care lifecycle:

• Prevention & Screening: S1 (Education via RAG/LLMs) and S2 (Opportunistic BMD estimation from CT).
• Diagnosis & Risk: S3 (Diagnostic reasoning), S4 (Fracture-risk stratification).
• Treatment Planning: S5 (Treatment-pathway reasoning with mandatory review), S6 (Physics-enhanced surgical planning for cement volume/material using finite-element analysis), and S7 (Intraoperative navigation and leakage risk prediction using fluid-dynamics surrogates).
• Long-term Management: S8 (Rehabilitation/adherence) and S9 (Longitudinal outcome monitoring).
  The workflow is deterministic: a routine CT triggers S2; if risk is high, parallel skills (S1, S3, S4) feed into S5 for a clinician-reviewed recommendation. If surgery is chosen, S6 and S7 assist in planning. Post-treatment, S9 monitors outcomes and can trigger a feedback loop to re-stratify risk (S4) if adverse events occur.

Key Contributions

1. Conceptual Abstraction: The paper defines "Clinical AI Skills" as the necessary abstraction between isolated models and future medical agents, embedding clinical meaning, safety constraints, and governance metadata directly into the capability unit.
2. Runtime Infrastructure: It introduces the "Clinical Harness," a governance architecture that converts abstract principles into executable, auditable, and interruptible system capabilities. It shifts the paradigm from "open-ended agent planning" to "deterministic clinical orchestration."
3. Hybrid Reasoning Integration: The framework demonstrates how knowledge-driven (guidelines), data-driven (ML models), and physics-enhanced (simulation/surrogates) skills can operate together within a single governed workflow.
4. Implementation Roadmap: The authors propose a four-phase validation pathway: skill-level validation, silent deployment (shadow mode), prospective workflow evaluation with clinicians in control, and post-deployment monitoring for drift and safety.

Results and Claims
The paper does not present empirical clinical trial results or performance metrics from a deployed system. Instead, it offers a conceptual and architectural proof-of-concept. The "results" are the structured definitions of the skill ecosystem and the logical demonstration of how such a system would function to manage osteoporosis care.

The authors claim that this framework addresses the organizational gap in AI deployment by:

• Ensuring accountability through deterministic orchestration rather than open-ended model planning.
• Extending AI from diagnostic intelligence to interventional intelligence (e.g., surgical planning) via physics-enhanced skills.
• Closing the loop on patient outcomes by feeding longitudinal data back into risk stratification.
• Providing a constrained execution substrate for future "governable medical agents," ensuring that agent autonomy is bounded by safety gates, audit trails, and human oversight.

Significance
The paper argues that the future value of medical AI depends not on the sophistication of individual models, but on the ability to organize validated models into safe, accountable clinical workflows. By reframing deployment as "governed capability organization," the Clinical Harness provides a reference architecture for transitioning from isolated predictions to lifecycle-spanning, AI-enabled care. The framework is presented as disease-agnostic, with potential applicability to other chronic conditions like heart failure and diabetes, ultimately serving as a bridge toward the safe integration of autonomous medical agents.