Life insurer improves underwriting efficiency with EXL XTRAKTO.AI™

Automated extraction of entities from medical documents to drive higher underwriting efficiency through improved mortality risk models


One of the leading life insurance companies in the U.S. had very large volumes of historical attending physician statement (APS) documents which they wanted to digitize. APS documents are extremely lengthy, completely unstructured, and contain complex medical language. These documents contain years of an applicant’s medical history including past diagnoses, medications, procedures, lab reports, and other information. It is highly resource intensive to manually go through these APS documents and extract information, as it requires people with specialized medical skills.

The client partnered with EXL for assistance extracting key medical impairments and features from APS documents to improve their mortality models and underwriting capability. They were looking for an AI and natural language processing (NLP) solution that could help them extract medical impairments and entities from these documents in an automated and quick way

A very large life insurance company leveraged EXL XTRAKTO.AI™ solution to accelerate key impairment identification from unstructured medical records, with extraction accuracy of up to 97%.


EXL assessed the client’s needs and recommended its EXL XTRAKTO.AI™ solution, an AI-powered document processing solution. EXL XTRAKTO. AI™ uses state-of-the-art AI techniques including machine learning, NLP, and computer vision. It is enriched with domain context through industry-specific language models and EXL’s domain ontology, including our healthcare and medical taxonomy. This medical ontology provides the solution with additional intelligence on medical conditions, their relationships , associated medications, and other domain specific business intelligence. Relevant pre-built components of EXL XTRAKTO.AI™ were configured together to deploy a solution capable of predicting impairments with high degree of accuracy. The overarching solution was also able to predict a set of over 185 feature for each APS document. These features include indicators for hierarchical condition categories (HCCs), diagnoses, ICD codes, and medications.

An intuitive user interface enables fast human validation of extracted information, leading to a low error and high quality of extraction output.

Additionally, the deployed solution is highly scalable, with the ability to process very large volumes at speed. The solution is also able to accommodate new use cases by leveraging pre-built components in an already established deployment software-as-a-service environment.


Key impairments identified with an accuracy range of 90-97%

~30 million pages of APS documents were processed in four weeks to provide over 185 medical features per document