Lift Off! How Health Actuaries Are Using Non-Traditional Features and Advanced Methods to Increase Predictability
While algorithms like XGBoost and recurrent neural networks are all the rave, the backbone of predictive analytics and machine learning comes down to the right data and a strong foundation of engineering that data. In this presentation, actuaries and data scientists from the Cleveland Clinic and Milliman discuss novel data points and features that have provided increased predictability in models used across health care. This data ranges from publicly available information to patient-entered questionnaires, and includes social determinants of health, third-party consumer information, measures of functional and emotional status, and new twists on familiar concepts. The group will also discuss methods in feature engineering and their applications to real-world problem solving. Data and feature engineering are what really provide the lift in predicting outcomes.
Hidden Bias - Identifying Unintended Consequences in Machine Learning
Over the past few years there have been several examples of machine learning models with unintended bias in applications ranging from hiring to provisioning healthcare services. As these models become more prevalent, identifying potential sources of bias and how to address them will become increasingly important. This session will cover some of the factors that can contribute to hidden bias as well as potential methods for identifying them. It will also address the role of actuarial subject matter expertise in combating potential issues.
Frontiers of Machine Learning Models for Actuarial Applications
Artificial Intelligence and machine learning models have grown in both sophistication and complexity and have demonstrated impressive accuracy and allowed deeper insights into claims, morbidity, customer stratification, and consumer behavior, to name a few. At the same time new techniques are being continuously developed, with ensemble learning methods replacing a single optimized model and new applications using Bayesian inference. To take advantage of this revolution in data, actuaries will need to understand both the theoretical underpinnings and practical applications of these models. We will briefly survey the landscape of models and explain how to construct and evaluate them, with an eye towards specific applications like group morbidity projection. We will also demonstrate concrete examples for how and why some fundamental models can be used to model uncertain quantities in health insurance in ways that improve on traditional actuarial predictions. Finally, we will look at the frontier of advanced analytics and survey cutting edge techniques such as Bayesian Additive Regression Trees that may be implemented in the future to improve actuarial practice.