AI and Predictive Analytics help reduce customer complaints by ~20% for a Health Insurance Provider

Case Study

AI and Predictive Analytics help reduce customer complaints by ~20% for a Health Insurance Provider

Business Objective

Our Client is a major US health insurance provider. One of their key plans, MAPD (Medicare Advantage Prescription Drug), caters to people aged 65 and older. The client wanted to identify dissatisfied members through ML frameworks to prevent them from filing grievances and CMS complaints through interventions. The reduced grievances and CMS complaints are expected to improve customer satisfaction & plan disenrollment. This should deliver a marked improvement in the plan’s Star rating making it attractive for prospects. More specifically, the client wanted to:

  • Develop a framework to predict members who are likely to file a first or repeat grievance
  • Identify top drivers that cause a member to file grievances or CMS complaints
  • Enable the client’s operations team to successfully intervene with the dissatisfied member

Challenges

  • The problem statement was relatively more unstructured
  • Highly imbalanced data- Grievance/Repeat grievance/Complaint rate is between 1- 3%

Solution Methodology 

  • Performed a comprehensive discovery exercise to define the problem:
    • Identified objective ways of measuring dissatisfaction, with an understanding that filing a grievance for the first time vs. repeatedly are two different behaviors
    • Evaluated different ways of sampling and designing the modeling construct
  • Built two separate models to predict the first and repeat grievance behaviour.
  • Developed specific hypotheses and tested them with the available data, through features related to claims, appeals, inquiries, network coverage, and customer profile
  • Performed several experiments and evaluated them iteratively to build the final model:
    • Tested feature values from different time periods (recent 30 days, 90 days, etc.)
    • Used various sampling techniques such as random under sampling, SMOTE
    • Used different modeling techniques such as Random Forest, GBM, XG Boost, MLP
  • Developed the final model, based on XGBoost, and cross-validated via Grid Search with a ~5x lift for the top decile.
  • Used SHAP technique to explain individual predictions of the machine learning models. Created a report to inform the operations team about the top drivers causing grievances for a customer, thereby enabling effective intervention strategies.

Business Impact

  • There is a projected estimate of 10% reduction in annual number of first grievances/complaints and an estimated 20% reduction in annual number of repeat grievances/complaints
  • Other business impact indicators include lower disenrollments from the plan, operational cost reduction, and improved Star ratings
Download Case Study

©2020 Tiger Analytics. All rights reserved.

Log in with your credentials

Forgot your details?