Commentary: Deep learning approaches applied to routinely collected health data: future directions.

Abstract

Prediction models have been a cornerstone of cardiovascular epidemiology for decades. Various types of methods have been tested for improvements in model performance, including machine learning models. In this issue of the International Journal of Epidemiology, Barbieri et al.1 combine survival methods with deep learning models to predict the 5-year risk of fatal or non-fatal cardiovascular events using nationally linked administrative databases. This study demonstrates two developments in health prediction research: the increasing use of linked administrative databases to generate predictive models and the application of deep learning methods to execute prediction tasks. The authors demonstrate that deep learning approaches can feasibly be applied to routinely collected administrative databases and suggest a performance advantage. What can we learn from studies that apply deep learning methods to health administrative data for prediction tasks? Moreover, what is needed to improve the application of deep learning methods for the prediction of cardiovascular and other health outcomes?