RESEARCH / INSURANCE
A staged approach to reading insurance statements and provider contracts: a layout model locates relevant regions, then a language model converts those regions into structured, auditable records.
Read on this pageInsurance documents contain tables, amounts and policy conditions whose meaning depends on where they appear on the page. This paper describes a staged extraction pipeline for Explanation of Benefits statements and provider contracts, designed to preserve that layout when producing structured records.
The pipeline first identifies the document type and relevant regions, then asks a language model to read those smaller regions. Its outputs bring claim lines, coverage information and service terms into tables that can be checked against the original document. The paper includes training and field-level evaluation details, showing where the method performs well and where longer policy language remains difficult. It is relevant to teams reducing manual document entry while retaining a path back to the evidence.
Layout understanding helps keep amounts and policy terms attached to the right context.
Focused image regions reduce the amount of material sent to the language model.
Field-level evaluation exposes harder free-text coverage conditions alongside simpler values.
Results are reported for the paper's EOB and provider-contract evaluation data. Extraction is imperfect, particularly for free-text policy fields. Performance on new layouts and downstream use of the records require validation rather than assuming every extracted field is correct.
The complete paper, with its original figures, tables and references.
20 pages · Read without leaving this pageDiscuss the approach, evidence and relevance to your team with the people building it.
