Research map / Vision-language and multimodal / Human and embodied modeling
KAMEL: Knowledge Aware Medical Entity Linkage to Automate Health Insurance Claims Processing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes KAMEL, a framework that links free-text health insurance underwriting exclusions to ICD-10CM codes by combining SBioBERT embedding similarity with ontology-based medical concept recognition and valid-ancestor linkage.
Problem
Automating health insurance claim rejection requires determining whether a claim's ICD-10CM diagnosis codes are related to free-text underwriting exclusions, a sparse, imbalanced, many-label problem where multi-label classifiers perform poorly on rare codes.
Approach
Combines implicit domain knowledge (a BioBERT-based Sentence-BERT encoder, SBioBERT, used for cosine similarity) with explicit domain knowledge (Disease Ontology plus a constructed three-level ICD ontology) to extract medical concepts from exclusions and link them to ICD codes via direct paths and learned valid ancestors, aggregating decisions with a MAX operator.
Key results
- KAMEL-DOICD reached F1 0.3924 vs 0.2352 (CAML) and 0.2593 (PLM-ICD) on Health Insurance I, with accuracy 0.9743.
- On MIMIC-IV-ICD-10-N3, KAMEL-DOICD achieved F1 0.7775 and accuracy 0.7750, while CAML and PLM-ICD scored F1 0.0000 and 0.1225.
- Adding explicit ontology inference to the implicit SSI module raised F1 on Health Insurance II from 0.5339 to 0.6423 (KAMEL-DOICD).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace CAML with KAMEL-DOICD | CAML · Health Insurance I | F1: 0.2352 -> 0.3924 | Helped |
| replace CAML with KAMEL-DOICD | CAML · Health Insurance II | F1: 0.3610 -> 0.6423 | Helped |
| replace PLM-ICD with KAMEL-DOICD | PLM-ICD · MIMIC-IV-ICD-10-N3 | F1: 0.1225 -> 0.7775 | Helped |
| add explicit domain knowledge inference to SSI (ablation) | KAMEL · Health Insurance I | F1: 0.1641 -> 0.3924 | Helped |
| add explicit domain knowledge inference to SSI (ablation) | KAMEL · Health Insurance II | F1: 0.5339 -> 0.6423 | Helped |
| add explicit domain knowledge inference to SSI (ablation) | KAMEL · MIMIC-IV-ICD-10-N3 | F1: 0.0000 -> 0.7775 | Helped |
| combine Disease Ontology with ICD Ontology (ablation) | KAMEL-ICD · Health Insurance I | F1: 0.2940 -> 0.3924 | Helped |
| combine Disease Ontology with ICD Ontology (ablation) | KAMEL-ICD · Health Insurance II | F1: 0.5685 -> 0.6423 | Helped |
| combine Disease Ontology with ICD Ontology (ablation) | KAMEL-ICD · MIMIC-IV-ICD-10-N3 | F1: 0.7384 -> 0.7775 | Helped |
Surprising findings
- The implicit-only SSI module scored F1 0.0000 on MIMIC-IV-ICD-10-N3 because whole discharge summaries add noise to embedding similarity.
- PLM-ICD achieved higher recall than KAMEL on both health insurance datasets but very low precision, yielding low F1.
Limitations
- Does not handle inclusion terms (e.g., 'except Cataract') in exclusions.
- Relies on completeness of medical ontologies; Disease Ontology lacks full ICD-10CM coverage.
- Private health insurance datasets are not released, limiting reproducibility.
Experimental setup
Datasets: Health Insurance I Health Insurance II MIMIC-IV-ICD-10-N3 Disease Ontology ICD-10CM ontology
Benchmarks: Health Insurance I Health Insurance II MIMIC-IV-ICD-10-N3
Metrics: accuracy F1 precision recall
Baselines: CAML PLM-ICD SSI (implicit-only ablation) KAMEL-ICD KAMEL-DO
Builds on: Sentence-BERT BioBERT Disease Ontology PLM-ICD Trove
Closest papers by meaning
- BoxLM: Unifying Structures and Semantics of Medical Concepts for Diagnosis Prediction in Healthcare (ICML 2025 · similarity 0.83)
- Multimodal Medical Code Tokenizer (ICML 2025 · similarity 0.81)
- Reasoning-Enhanced Healthcare Predictions with Knowledge Graph Community Retrieval (ICLR 2025 · similarity 0.81)
- DearLLM: Enhancing Personalized Healthcare via Large Language Models-Deduced Feature Correlations (AAAI 2025 · similarity 0.80)
- SM3-Text-to-Query: Synthetic Multi-Model Medical Text-to-Query Benchmark (NeurIPS 2024 · similarity 0.80)
- Enhancing Small Medical Learners with Privacy-preserving Contextual Prompting (ICLR 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.