An Adaptable LSTM Network Predicting COVID-19 Occurrence Using Time Series Data

Abstract

As the COVID-19 pandemic progresses, it has become critical for policymakers and medical officials to understand how cases are trending. Machine learning models, particularly deep learning LSTM (Long Short-Term Memory) models, may hold immense value to forecast changes in COVID-19 cases. In this paper, a novel LSTM-based architecture is proposed, developed and trained on human logistics data that includes travel patterns, visits to commercial properties, as well as historical cases, demographic, and climate data. This data includes both time series and static data allowing the LSTM to be used in both classification and regression tasks to predict COVID-19 occurrence trends. For classification, the problem is modeled as a multiclass supervised learning classification problem with varying granularity. The proposed LSTM network achieves an 81.0% F1-score outperforming conventional machine learning model benchmarks (such as the random forest model with an F1 score of 58.9 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> ) and is comparable in performance to a time series forest model. Additionally, the LSTM model is adaptable to perform regression and predict a 14-day sliding window based on currently observed data with a mean absolute error of 0.0026. This research serves as a foundation for future work in the forecasting of COVID-19 and other similar disease outbreaks using similar temporal and static data.