Our client is a leading US-based provider of insurance and financial services products including life insurance, retirement plans, and mutual funds.
The underwriting team had a significant volume of informal underwriting inquiries via APS records. The client wanted to automate the process of analyzing unstructured APS records to enable straight-through processing for quote generation.
- Attending physician statements are unstructured and difficult to process manually.
- Reconciling and deriving KPIs definition from the APS records needed significant domain inputs from medical directors as well as detailed understanding of the underwriting process
- To confirm to audit standards, white box interpretable models were needed for predicting risk classes
- Cleaned and created derived KPIs based on the given medical context
- Leveraged advanced text mining techniques to extract features around health, vitals, treatment history, drug patterns and lifestyle preferences from unstructured APS summaries
- Created vector encoding using TF-IDF from domain-specific APS text and narrowed down to specific vectors that were significantly impacting risk behavior
- Designed an end-to-end underwriting framework and built a predictive model to assess/classify risk associated with submitted records
- Furthermore, designed knock out rules to improve risk prediction accuracy to account for low-frequency high-risk patterns which were weak signals in the model
- ~40% of quotes processed straight through leading to significant efficiency gains
- Faster APS Submit-to-Quote time
- End-to-end interactive process flow leading to significant cost reduction.