Machine Learning for National Mortality Forecasting Using Extreme Heat Events

A recent paper utilized a deep learning methodology when analyzing multivariate time series data to forecast mortality in Southern Africa. However, the high distributional variance of the data hindered the performance of the model. Another recent paper utilized a statistical machine learning approach to analyze mortality in the United States but failed to provide a comprehensive forecast as parts of the dataset were removed. Our paper seeks to improve upon these previous methods by utilizing Extreme Heat Events (EHE) data and an autoregressive machine learning model for forecasting mortality. We are able to demonstrate a high correlation between EHE's and Mortality through their higher accuracy scores in forecasting national mortality through two comparative experiments. Additionally, the high 0.91 R Squared accuracy score of our autoregressive multilayer perceptron model further reinforced the necessity for the role of machine learning in mortality forecasting. Ultimately, our research demonstrates the importance of including EHE's and machine learning methods as factors when considering mortality forecasting.