Payer-Side Medical Audit for SOAT Claims in Colombia: A Methodological Description

Technical Summary: Payer-Side Medical Audit for SOAT Claims in Colombia

Problem Statement
Medical claims review in health systems is a critical administrative and clinical function where payers must verify that submitted claims are complete, clinically justified, and compliant with regulatory and contractual requirements. In Latin American health systems, and specifically in Colombia, this process is complicated by fragmented documentation, heterogeneous provider workflows, and evolving digital billing mandates. The Colombian context adds specific regulatory layers, including the Seguro Obligatorio de Accidentes de Tránsito (SOAT) for traffic accidents, the Administradora de los Recursos del Sistema General de Seguridad Social en Salud (ADRES) reimbursement workflows, and strict documentation standards (e.g., FURIPS, RIPS/FEV, and clinical records).

Existing literature on AI-assisted billing often focuses on fraud detection or revenue-cycle management in high-income settings, utilizing anomaly detection or supervised learning. However, these approaches frequently fail to address the specific methodological needs of payer-side audit in Latin America: preserving documentary traceability, distinguishing between "return" (missing documents) and "objection" (clinical/regulatory non-compliance), and consolidating a provider-facing ruling that is legally defensible. There is a gap in published methodologies that describe how to separate documentary, clinical, and ruling logic while maintaining mandatory human oversight and source-level traceability in a fragmented regulatory environment.

Methodology
This study presents a methodological description of a six-phase, rule-grounded payer-side audit pipeline designed for SOAT/ADRES claims in Colombia. The system operates within the "Salmona" operational environment developed by Arkangel AI.

• Architecture and Design: The pipeline is structured as a sequence of versioned, distinct instruments rather than a single monolithic prompt. It utilizes OpenAI Codex (via API) in a reasoning mode with fixed prompting and predefined JSON output schemas to ensure procedural consistency. The system does not use temperature parameters, relying instead on structured constraints.
• The Six-Phase Workflow:
  1. Phase 1a (Support Validation): Verifies the presence of 21 expected document types (12 mandatory, 9 conditional). A "return" verdict here terminates the process, indicating the claim is not formally reviewable.
  2. Phase 1b (Data Structuring): Maps the claim file into 62 standardized fields, anchoring each to a specific document, page, and section. It maintains a financial invariant where approved plus objected values must reconcile to the total invoice.
  3. Phase 1c (Cross-Document Consistency): Applies seven checks (identity, event date, provider ID, vehicle data, totals, diagnoses, surgical breakdown) to identify contradictions. Blocking inconsistencies trigger an early exit.
  4. Phase 2a (Clinical Audit): Runs concurrently with Phase 2b. It evaluates 25 rules regarding diagnosis-procedure coherence, clinical pertinence, and payer business logic. Rules generate textual evidence and confidence levels.
  5. Phase 2b (Atypicality Detection): Evaluates 14 patterns against historical or statistical baselines. Crucially, findings here do not directly alter financial decisions; they serve as flags for investigation.
  6. Phase 3 (Final Ruling): Consolidates all executed phases, applies six prioritized decision rules (including a specific subpath for documentary returns), re-verifies monetary reconciliation, and generates the formal provider letter.
• Control Mechanisms:
  • Early-Exit Gates: Two gates route cases directly to the Final Ruling phase if documentary incompleteness (Gate 1) or blocking inconsistencies (Gate 2) are detected, preventing clinical audit on unreviewable files.
  • Human Review: A mandatory checkpoint exists. If the confidence for a clinical rule falls below 75.0%, the case is flagged for human review. No ruling is communicated without human supervision.
  • Traceability: Every extracted field and rule output is linked back to the source document, page, and section.

Key Contributions
The paper contributes a structured framework for payer-side audit that addresses the specific regulatory and operational constraints of the Colombian health system.

• Separation of Concerns: The architecture explicitly decouples documentary validation, data structuring, clinical audit, atypicality detection, and final ruling into independent instruments. This prevents the conflation of evidence extraction, clinical judgment, and fraud suspicion.
• Regulatory Encoding: The system encodes specific Colombian regulations (e.g., Decreto 056 de 2015, Resolución 3047 de 2008) into versioned instruments, allowing for regulatory updates without rewriting core pipeline logic.
• Traceability and Safety: By preserving source-level references and enforcing a "conservative" workflow (where incomplete data leads to a return rather than a clinical objection), the system ensures that audit findings are grounded in readable, consistent evidence.
• Differentiation from Provider-Side Audit: Unlike the provider-side pipeline (which uses parallel layers), the payer-side module uses a sequential flow with early exits, reflecting the payer's need to reject incomplete packages before clinical review.

Results
The study provides a functional demonstration rather than a statistical validation against an expert-auditor reference standard.

• Case Selection: 15 de-identified cases were selected from operational runs to cover diverse claim profiles.
• Coverage: The demonstration successfully executed end-to-end across all six decision rules (D01–D06) and both early-exit gates.
  • 5 cases resulted in total approval.
  • 6 cases resulted in partial objection.
  • 3 cases resulted in return (1 via Gate 1, 2 via Gate 2).
  • 1 case was escalated to human review due to illegible documents.
• Financial Reconciliation: For the 14 closed cases (total billed: COP 137.7 million), the system verified that the sum of approved and objected/returned amounts matched the billed total.
• Limitations of Evidence: The authors explicitly state that no metrics such as precision, recall, F1 score, or inter-auditor agreement were computed. The cases serve to demonstrate execution capability, not to validate accuracy against a gold standard.

Significance and Claims
The paper modestly claims to fill a methodological gap in the literature regarding payer-side audit in Latin American health systems. Its significance lies in providing a replicable framework that:

1. Preserves Human Oversight: It positions AI as an assistive tool that structures inputs and applies rules but requires mandatory human review for final decisions, particularly when confidence is low.
2. Ensures Legal Defensibility: By maintaining a clear audit trail from source document to ruling, the system supports the administrative and regulatory requirements of Colombian health law.
3. Adapts to Fragmentation: The phase-separated design allows the system to handle the fragmented documentation and local regulatory complexity characteristic of the region, which general-purpose automation often fails to address.

The authors conclude that prospective validation against expert-auditor decisions is required to establish the operational value, accuracy, and safety of the system. The current work establishes the architectural feasibility and end-to-end execution of a rule-grounded, traceable audit pipeline.