Predicting Healthcare Costs

Overview

I created the user interface for a project whose aim was to develop machine learning models for a health insurance company that could out-perform its existing actuarial models in predicting the annual per-member costs of subscribers. The purpose of the user interface was to engender trust in the results of the models and compare the performance of different models.

Summary

Early in the year, when prices needed to be set, actuarial models were not very good at predicting the cost of healthcare. However, the company was apprehensive about adopting machine learning models whose drivers could not be as clearly explained as those in actuarial models, even if the results were more accurate. The goal of this interface was to allow a comparison over time between the predictions of the different machine learning models and the actual costs and to show which features most influenced the results of a model alongside the factors that drove the actual costs. These metrics and comparisons ideally allow the business user to determine at what point a given model has been consistent enough to trust, which model was the best to use, and to explore the relationship between the model drivers and the actual drivers.