Agentic AI for Healthcare Claim Denial Analysis: A Policy-Grounded Framework for Privacy Preservation

Sofia Campos, Carlos Martins, Pedro Pires, João Gama, Bruno Veloso, Pedro SantosarXiv 2026 · arXiv

This paper presents a modular, multi-agent AI framework designed to tackle one of the most persistent headaches in the healthcare industry: claim denials. By combining predictive analytics, graph-based learning, and a policy-grounded Retrieval-Augmented Generation (RAG) pipeline, the authors demonstrate a system that not only predicts denials but explains them through the lens of specific insurance policies while strictly adhering to HIPAA and GDPR privacy standards.

TL;DR

The authors introduce an "Agentic AI" framework that uses specialized sub-agents and policy-grounded RAG to analyze healthcare claim denials with high transparency. Its standout feature is a dedicated HIPAA Compliance Agent that successfully redacted 100% of sensitive identifiers in testing while maintaining 98% traceability of its analytical findings.

What Problem Does This Paper Address?

Healthcare claim denials—where insurers refuse to pay for medical services—impose a massive administrative and financial burden on providers. While AI offers a path toward automation, the sector faces three major hurdles:

  1. Explainability: Standard "black-box" models can't tell a billing specialist why a claim was denied or how to fix it.
  2. Policy Grounding: Denials are often tied to complex, frequently changing payer policies that general-purpose LLMs don't know.
  3. Privacy: Handling Protected Health Information (PHI) requires strict adherence to regulations like HIPAA and GDPR, making it risky to send raw data to cloud-based LLMs.

The authors aim to bridge the gap between high-performance predictive modeling and the rigorous regulatory requirements of the healthcare domain.

Methods

The framework utilizes a top-down hierarchical architecture orchestrated by a Coordinator Agent (powered by gpt-oss-120b). As shown in the workflow below, the Coordinator delegates tasks to four specialized sub-agents:

Fig. 1: Agentic AI workflow for healthcare claim-denial analysis: a Coordinator Agent delegates to four specialized sub-agents (Claim Denial Predictor, Risk Drivers, Code Combinations, Counterfactual Causality), enriched by a Hybrid Contextual RAG pipeline and finalized by a HIPAA/GDPR Compliance Agent.
  • Claim Denial Predictor: Uses an Optuna-tuned XGBoost model to estimate denial probability.
  • Risk Drivers Agent: Employs TreeSHAP to identify which specific features (e.g., processing time, claim amount) are pushing the claim toward rejection.
  • Code Combinations Agent: Uses Graph Neural Networks (GNNs) to find high-risk relationships between diagnosis, procedure, and medication codes.
  • Counterfactual Causality Agent: Performs "what-if" analysis to suggest actionable changes (like obtaining pre-authorization) that could flip a predicted denial to an approval.

The system is grounded by a Hybrid Contextual RAG pipeline that retrieves relevant payer policies using a mix of BM25 (lexical) and dense semantic embeddings. Finally, a HIPAA/GDPR Compliance Agent performs deterministic redaction of PHI and an LLM-based review to ensure the final report is safe for distribution.

The specific inputs and outputs for these agents are summarized in the table below:

Table 1: Overview of the specialized agents: tools, inputs, and outputs.

Datasets

Because real-world healthcare data is highly sensitive and difficult to share, the authors developed and evaluated their framework on a synthetic healthcare claims dataset.

  • Volume: 99,079 adjudicated claims.
  • Realism: It reproduces the structure of U.S. hospital billing systems, including ICD-10 codes, payer names, and administrative life-cycles, but contains no real patient information.
  • Split: A stratified 80/20 train-test split was used across all models.

Key Insights and Ideas

  • Agent-as-Tool Paradigm: Rather than relying on an LLM to "guess" denial risks, the framework treats specialized ML models (XGBoost, GNNs) as tools. This combines the reasoning capabilities of LLMs with the statistical rigor of traditional machine learning.
  • Contextual RAG: The authors found that simply chunking policy documents wasn't enough. By "enriching" each chunk with LLM-generated context (identifying the specific payer and benefit), they significantly improved retrieval coherence.
  • Privacy as a "Quality Gate": The HIPAA Compliance Agent acts as a downstream filter. It doesn't interfere with the analysis; it ensures the presentation of that analysis is compliant. This allows for complex reasoning on raw data within a secure environment while outputting safe, de-identified reports.
  • Actionable Counterfactuals: One of the most useful outputs is the "recourse" analysis. Telling a user that a claim has an 80% risk of denial is helpful; telling them that reducing processing time to the typical level would drop that risk to 7% is transformative.

Key Results

The framework was evaluated on 100 held-out high-risk claims with impressive results:

  • Faithfulness: 98% of numerical values in the generated reports were successfully traced back to the underlying tool evidence.
  • Citation Accuracy: The RAG system validated 214 out of 216 policy citations, with zero "hallucinated" documents.
  • Privacy Preservation: The compliance layer identified and removed all 280 sensitive identifiers (payer names, claim IDs, PHI patterns) found in the draft reports.
  • Efficiency: A complete, multi-agent analysis took an average of 17–18 seconds per claim.

Limitations and Future Work

Despite the success, the authors noted several areas for improvement:

  • Technical Jargon: Only 23% of the reports were entirely free of technical jargon (like "SHAP" or "lift"). Future iterations need better deterministic enforcement of "plain language" summaries for non-technical billing staff.
  • Synthetic vs. Real Data: While the synthetic data is robust, the framework needs validation on real-world, messy clinical data to prove its generalizability.
  • Architectural Scaling: The authors suggest investigating Agent-to-Agent (A2A) communication and the Model Context Protocol (MCP) to improve how agents share data and interoperate with external systems.

Summary updated

Abstract

In the healthcare sector, claim denials impose significant financial and administrative burdens, emphasizing the need for automated and trustworthy decision-support systems. This paper proposes a modular Agentic Artificial Intelligence (AI) framework for healthcare claim denial analysis, integrating specialized agents for denial risk prediction, risk factor identification, diagnosis-procedure-medication code relationship analysis, and counterfactual analysis. The framework integrates predictive analytics, graph-based learning, a Hybrid Contextual Retrieval-Augmented Generation (RAG) pipeline grounded in healthcare policy documents, and a dedicated HIPAA Compliance Agent to support privacy protection. Evaluated on a synthetic healthcare claims dataset, the framework achieved 98% traceability of reported numerical findings, validated 214 of 216 policy citations, and removed all 280 identified sensitive identifiers without modifying the underlying analysis. These findings pave the way for further research into Agentic AI as a framework for explainable, policy-aware, and privacy-preserving healthcare claim denial management.
BibTeX
@misc{campos2026agentic,
  author = {Sofia Campos and Carlos Martins and Pedro Pires and João Gama and Bruno Veloso and Pedro Santos},
  title = {{Agentic AI for Healthcare Claim Denial Analysis: A Policy-Grounded Framework for Privacy Preservation}},
  year = {2026},
  eprint = {2611.09423},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.09423},
}
Figure