Introduction
In the health insurance industry, achieving Straight-Through Processing (STP)—the seamless, automated adjudication of claims—remains a significant challenge. A primary obstacle is the interpretation of "Underwriting (UW) Exclusions." These are policy-specific rules written in natural language that exclude coverage for certain conditions based on a member's medical history. Bridging the gap between these free-text descriptions and the standardized ICD-10CM codes used in hospital billing is a complex task that traditional machine learning models often fail to solve effectively.
This paper introduces KAMEL (Knowledge Aware Medical Entity Linkage), a framework that integrates pre-trained language models with structured medical ontologies to automate this linkage. By leveraging both implicit semantic understanding and explicit hierarchical logic, KAMEL achieves superior performance, particularly in handling rare medical conditions where data is sparse.
The Challenge of Underwriting Exclusions
Underwriting exclusions are typically broad and descriptive, such as "Excluded from any disease or disorder of either eye." Conversely, medical claims are submitted using highly specific ICD-10CM codes, such as H25 (Age-related cataract) or H40 (Glaucoma). With over 69,000 possible codes, mapping a single natural language phrase to all relevant diagnostic codes is a daunting "many-to-many" matching problem.
Standard AI models struggle with "data sparsity"—they cannot accurately predict codes they haven't encountered frequently in training data. Furthermore, they lack the inherent medical logic to understand that a "cataract" is a subset of "eye diseases" unless explicitly taught.
The KAMEL Framework
KAMEL addresses these issues through a hybrid approach that combines Implicit and Explicit domain knowledge.
1. Implicit Knowledge (SBioBERT)
The system utilizes a Siamese BioBERT (SBioBERT) model to generate semantic embeddings of both the exclusion text and the ICD descriptions. This allows the model to recognize that terms like "cardiovascular" and "heart" are related, even without an exact word match.
2. Explicit Knowledge (Medical Ontologies)
KAMEL incorporates the Disease Ontology (DO) and a custom ICD Ontology. These structured knowledge bases act as a map, defining the hierarchical relationships between medical concepts. For example, the ontology identifies that "Lens Disease" is a child of "Eye Disease," and "Cataract" is a child of "Lens Disease."
3. Medical Concept Recognition (MCR) and Valid Ancestors
Using a sliding window approach, the MCR module identifies medical terms within the free-text exclusions and maps them to nodes in the ontologies. The core innovation of KAMEL is the use of Valid Ancestors (VAs). Instead of searching for exact matches, the system identifies common points in the ontology tree that connect a policy exclusion to a claim's diagnosis. If a claim's ICD code and a policy exclusion share a relevant ancestor in the hierarchy, the claim is flagged for potential rejection. Finally, KAMEL uses a "MAX" operator to aggregate results from both global sentence-level similarity and local concept-level matching, ensuring comprehensive coverage.
Experimental Setup and Datasets
The framework was rigorously evaluated across three distinct datasets:
- Health Insurance I & II: Large-scale, proprietary datasets from international insurance providers, containing real-world claims and policy exclusions.
- MIMIC-IV-ICD-10-N3: A public dataset derived from the MIMIC-IV clinical database, adapted to simulate the insurance exclusion detection task.
Results and Key Insights
KAMEL significantly outperformed state-of-the-art multi-label classification benchmarks, including CAML and PLM-ICD.
- Superior Performance on Public Data: On the MIMIC dataset, KAMEL achieved an F1-score of 0.7775. In contrast, CAML failed to produce results due to the complexity of discharge summaries, and PLM-ICD reached only 0.1225.
- Real-World Robustness: On the "Health Insurance I" dataset, KAMEL achieved an accuracy of 97.43%, demonstrating its readiness for production environments.
- Explainability: Unlike "black-box" neural networks, KAMEL provides a transparent logic trail (e.g., Eye Disease -> Lens Disease -> Cataract). This transparency is crucial for regulatory compliance and for providing clear justifications to policyholders when a claim is rejected.
Conclusion and Future Work
KAMEL demonstrates that a hybrid approach—combining the contextual power of language models with the structured logic of medical ontologies—is the key to automating complex insurance workflows. While highly effective, the authors note that future iterations will need to better handle complex linguistic logic, such as negations (e.g., "Exclude all eye conditions except cataracts"). Expanding the ontologies to include more diverse disease-to-disease relationships will further enhance the system's reasoning capabilities.
Summary updated