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 Santos

cs.AINov 6, 2026

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

Authors: Sofia Campos, Carlos Martins, Pedro Pires, João Gama, Bruno Veloso, Pedro Santos

Organizations: Randy Labs, Portugal. · University of Porto, Portugal. · INESC TEC, Portugal.

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.

Figures & tables

Figure 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.

Agent / Tool	Input	Structured output
Claim Denial Prediction get_prediction	Calibrated denial probability from the Optuna-tuned XGBoost model; test-set reliability metrics; actual outcome (if known)	Denial probability, predicted outcome vs. decision threshold, risk level, and model reliability in plain language
Risk Drivers get_shap_explanation get_population_statistics	Per-claim TreeSHAP contributions as plain-language drivers with strength; claim attributes and population percentiles; payer denial rate	Drivers toward rejection and toward payment (strong / moderate / small) and comparison with the population
Code Combinations get_code_combinations	Code co-occurrence graph with GNN code propensities; pair/triple combination lift; claim-level GNN probability; code descriptions	High-risk code combinations (labelled codes, denial rate, $\times$ average, support), riskiest single codes, and a HIGH / MEDIUM / LOW verdict
Counterfactual Causality get_counterfactual_recourse	Claim re-scored by the calibrated XGBoost model under actionable interventions (approved pre-authorization, typical processing time)	Probability before $\rightarrow$ after each intervention, decision flips, best achievable outcome, and most effective lever
HIPAA / GDPR Compliance	Coordinator’s draft report; HIPAA passages retrieved from a dedicated index	Redacted report (PHI, PAYER-xxxxxx , CLAIM-xxxxxx ) and a JSON verdict: HIPAA / GDPR compliance, payer anonymity, issues, and citations

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

Aspect	Description
Scope	Synthetic adjudicated healthcare claims associated with simulated clinical encounters from January to April 2026.
Volume	99,079 eligible claims retained from 100,000 generated records containing 81 columns. Eligibility requires a final claim status and a billed amount greater than $10. Each record is identified by a claim_id .
Target Variable	Binary classification: Denied/Partially Paid (Class 1): 27,183 claims (27.4%). Paid (Class 0): 71,896 claims (72.6%).
Data Realism	Real coding systems, including ICD-10 diagnosis codes, procedure, and medication codes with descriptions. Payer names, patient history and identifiers, are fictional.
Feature Composition	Claim attributes (amount, type, setting, month, and day of week); administrative variables (processing time and pre-authorization status); payer and patient history (average amounts, claim counts, and recent denial ratio); demographics (age band and gender); and clinical codes (diagnoses, procedures, and medications), comprising 143 distinct codes.
Data Split	A single stratified 80/20 train–test split, comprising 79,263 training claims and 19,816 test claims, shared across all evaluated models.

Table 2: Overview of the Synthetic Healthcare Claim Denial Dataset.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Type	Search space	Selected
Ensemble and tree structure
n_estimators	integer	$\{200,300,\ldots,600\}$	400
max_depth	integer	$\{3,4,\ldots,7\}$	6
min_child_weight	integer	$\{1,2,\ldots,5\}$	2
Learning rate and sampling
learning_rate	continuous (log)	$[0.005,0.2]$	0.0119

Appendix

Table A1: XGBoost hyperparameter search space explored with Optuna and the selected configuration.

Metric	Value	95% CI / uncertainty	Denominator / scope
Planning
Coordinator-selected mandatory steps	100%	96.3–100% a	100 claims
Claims requiring quality-gate completion	0/100	$\leq 3\%$ b	100 claims
Evidence
Mean numerical agreement per report	98%	not computed c	100 reports
Retrieval

Appendix

Table A3: Evaluation of the agentic framework across planning, evidence, retrieval, privacy, compliance, readability, prediction, recourse and cost dimensions.

AI Agent Architecture

See all 36 papers

Agentic Governance and Adversarial Verification for Policy-Constrained LLM Healthcare Appeal Generation
Aug 20, 2026Harshil Lodhiya, Alex McManus, Reese WalkerAI Agent Governance Multi-Agent LLM Systems
Sliced Health, Woodstock, GA, United States
From Siloed Algorithms to Compliance-First Agentic Platforms: A Multi-Layered Architecture for Hospital AI Systems
Aug 6, 2026Manideep Dhar, Ritwik Singh, Sharat Chandra Kumar ManikondaMulti-Agent Orchestration Healthcare
Chief Operating Officer and Head of Research and Development, Instil-IT, Hyderabad, Telangana. · Chief Artificial Intelligence Architect Engineer, Instil-IT, Hyderabad, Telangana. · Chief Executive Officer, Managing Director and Chief Data Scientist, Instil-IT, Hyderabad, Telangana.
Semantic Risk Scoring of Aggregated Metrics: An AI-Driven Approach for Healthcare Data Governance
Mar 9, 2026Mohammed Omer Shakeel AhmedPrivacy Auditing Data Leakage
University of Texas at Arlington · Dallas, TX

Aug 20, 2026J/K move · Enter open · S save

Agentic Governance and Adversarial Verification for Policy-Constrained LLM Healthcare Appeal Generation

Harshil Lodhiya, Alex McManus, Reese Walker

Sliced Health, Woodstock, GA, United States

AI Agent Governance Multi-Agent LLM Systems LLM Agent Verification Medical LLMs Controllable Text Generation

Open paper AskPDF arXiv

Claim denial management costs U.S. healthcare approximately $260 billion annually in administrative overhead. Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) can produce fluent clinical text, but single-agent architectures fail in high-stakes healthcare: they introduce unsupported clinical details and lose the logical structure of hierarchical payer policy. We propose AGVF (Agentic Governance and Adversarial Verification Framework), a multi-agent architecture for medical-necessity appeal generation under explicit policy and evidence constraints. AGVF models appeal synthesis as a Constrained Markov Decision Process (CMDP) over five agents: policy formalization, evidence retrieval, gap analysis, adversarial critique, and gated synthesis. We prove that refinement over a fixed policy constraint graph monotonically reduces evidence-deficiency and terminates with either a complete satisfying frontier or a localized evidence gap. A deterministic citation- grounding gate prevents assertions without admissible evidence from entering shared state. We provide a reference implementation and validate it on 1,000 synthetic appeal cases parameterized from de-identified public hospital discharge data. The validation confirms zero citation-grounding violations across all AGVF cases and monotone deficiency reduction in every episode; ablating the gate raises violations to 100%, confirming it is load-bearing. The study uses no real patient records and does not measure clinical efficacy. AGVF thus contributes a theory-backed agentic architecture and verified reference implementation for policy-constrained LLM generation in healthcare.

Abstract

AI Agent Architecture

Explore similar work

Agentic Governance and Adversarial Verification for Policy-Constrained LLM Healthcare Appeal Generation

From Siloed Algorithms to Compliance-First Agentic Platforms: A Multi-Layered Architecture for Hospital AI Systems

Semantic Risk Scoring of Aggregated Metrics: An AI-Driven Approach for Healthcare Data Governance