“The client leveraged EXL’s AI- and NLP-driven solution, EXL Xtrakto.AI™, to automate identifying impairments in the APS documents in an automated manner.”
One of the largest life insurance companies in the US had large volume of historical attending physician statements (APS). APS documents are lengthy and unstructured. These statements contain many sections, such as a patient’s medications, diagnoses, and lab reports, often spread across multiple years of an applicant’s medical history. Additionally, APS documents contain complex medical terminology, which would require natural language processing (NLP) capabilities capable of understanding its nuances.
The client had two objectives. First, the client wanted to extract structured information from these unstructured medical records. This information could then be analyzed in several ways, leading to outcomes including improved underwriting. Second, the client wanted to identify impairments from APS documents using AI and NLP in an automated, accurate manner. This could potentially save time and effort for both clinical and underwriting experts by automating this traditionally manual process.
Human Ingenuity in Action
The client leveraged EXL’s AI- and NLP-driven solution, Xtrakto.AI™, to automate identifying impairments in the APS documents in an automated manner. The solution, backed by machine learning and NLP capabilities, was tuned to fit clients’ requirements. The NLP features were engineered to leverage a medical ontology database and healthcare domain knowledge to understand the complex medical terminology present in the APS documents. A variety of traditional solutions were implemented as well, such as word embedding-based NLP models trained with the goal of predicting impairments with high accuracy.
“With the client’s collaboration, EXL’s Xtrakto.AI™ solution achieved a high 90-97% accuracy rate for identifying impairments.”
With the client’s collaboration, EXL’s Xtrakto.AI™ solution achieved a high 90-97% accuracy rate for identifying impairments. AI and NLP models also provided evidence corresponding to predictions by returning the relevant text chunks in APS that indicated the presence of impairments .The AI and NLP solution developed is a highly scalable, and can be extended for many more impairments quickly without significant manual intervention. Additionally, this solution could also be repurposed for other NLP tasks on medical records.