Predictive Models for Decision Support in the COVID-19 Crisis

The task known as prediction is widely applied in several different areas of knowledge, from popular applications such as weather forecasting, going through supply chain management, an increasing range of adoption in healthcare and, more specifically in epidemiology, the central topic of this book. The new challenges brought with the COVID-19 pandemic highlighted the possibilities and necessity of using prediction techniques to support decisions related to epidemiology in both managerial and clinical areas. In practice, the current outbreak created a strong need for the adoption of different computational models to support both medical teams and public health administrators. The methods vary from simple linear regressions to very complex algorithms based on Artificial Intelligence (AI) techniques. The present chapter contextualizes the use of prediction for decision support as a foundation of the following chapters which are focused on the application for the COVID-19 pandemic time series. With such a large number of methods for data-driven predictions, a clear distinction between explanation and prediction is firstly provided. From there, a methodological framework is presented, from the data source definition and selection of countries as references for the analysis, going through data handling for validation, until the definition of the evaluation criteria for the proposed models.

The process of decision-making when dealing with infectious diseases is firmly based on mathematical modeling nowadays. One usual approach is to consider the adoption of compartmental methods such as SIR and SEIR and a large number of corresponding variations for modeling and prediction epidemic time series. Nevertheless, the COVID-19 epidemic characteristics and curves are apparently challenging the results obtained by these models. This chapter presents the results of two traditional compartmental models, SIR (Susceptible—Infected–Recovered) and SEIR (Susceptible–Exposed–Infected–Recovered), and an adapted version of the SEIR, called SEIR with Intervention, which captures the impact of containment measures for the dynamics of the infection rate. The analysis is performed for five countries: China, United States, Brazil, Italy, and Singapore, each of them with specific characteristics of dealing with the pandemic. A sequence of results is presented, considering different parameters, in order to understand the feasibility of application for each model.

When considering time-series forecasting, the application of autoregressive models is a popular and simple technique that is usually considered. In this chapter, we present the basic theoretical aspects and assumptions of the ARIMA—Autoregressive Integrated Moving Average model. It is considered for the prediction of the COVID-19 epidemiological data series of five different countries (China, United States, Brazil, Italy, and Singapore), each of them with specific curves, which are results of the virus reproduction itself but also of policies and government decisions during the pandemic spread. The discussion about the results is performed with the focus on the three evaluation criteria of the model: \(R^2\) Score, MAE, and MSE. Higher \(R^2\) Score was obtained when the sample time series was smoothly increasing or decreasing. The error metrics were higher when the prediction was performed for oscillating data series. This may indicate that the use of ARIMA models may be suitable as a prediction tool for the COVID-19 when the country is not facing severe oscillations in the number of infections.

Considering the application of prediction techniques to support the decision-making process during a dynamic environment such as the one faced during the COVID-19 pandemic, demands the evaluation of several different strategies to compare and define the most suitable solution for each necessity of prediction. Analyzing the epidemic time series, for example, the number of new confirmed cases of COVID-19 per day, classic compartmental models or linear regressions may not provide results with enough precision to support managerial or clinical decisions. The application of nonlinear models is an alternative to improve the performance of these models. The Kalman Filter (KF) is a state-space model that is used in several applications as a predictor. The filter algorithm requires low computational power and provides estimates of some unknown variables given the measurements observed over time. In this chapter, the KF predictor is considered in the analysis of five countries (China, United States, Brazil, Italy, and Singapore). Similarly to the ARIMA methodology, the results are evaluated based on three criteria: \(R^2\) Score, MAE (Mean Absolute Error), and MSE (Mean Square Error). It is important to notice that the definition of a predictor for epidemiological time series shall be carefully evaluated and more complex implementations do not always represent a better prediction on average. For the proposed KF predictor, there were specific time-series samples with no satisfactory result, achieving a negative \(R^2\) Score, for example, while, on the other, other samples achieved higher \(R^2\) Score and lower MAE and MSE, when compared to other linear predictors.

The use of computational intelligence techniques is being considered for a vast number of applications not only because of its increasing popularity but also because the results achieve good performance and are promising to keep improving. In this chapter, we present the basic theoretical aspects and assumptions of the LSTM model and H20 AutoML framework. It is evaluated on the prediction of the COVID-19 epidemiological data series for five different countries (China, United States, Brazil, Italy, and Singapore), each of them with specific curves, which are results of policies and decisions during the pandemic spread. The discussion about the results is performed with the focus on three evaluation criteria: \(R^2\) Score, MAE, and MSE. Higher \(R^2\) Score was obtained when the sample time series was smoothly increasing or decreasing. The results obtained by the AutoML framework achieved a higher \(R^2\) Score and lower MAE and MSE when compared with LSTM and also with other techniques proposed in the book, such as ARIMA and Kalman predictor. The application of machine learning algorithm selector might be a promising candidate for a good predictor for epidemic time series.

The support provided by geographic data and the corresponding processing tools can play an essential role to support decision-making process, especially for public healthcare during the current pandemic outbreak of the COVID-19. Geographic data collection may be challenging when is necessary to obtain precise latitude and longitude, for example. The current chapter presents a new tool for the geographic location prediction of new cases of COVID-19, considering the confirmed cases in the city of Fortaleza, capital of the State of Ceara, Brazil. The methodology is based on a sequential approach of four clustering algorithms: Agglomerative Clustering, DBSCAN, Mean Shift, and K-Means followed by a two-dimensional predictor based on the Kalman filter. The results are presented following a case study approach with different examples of implementation and the corresponding analysis of the results. The proposed technique could generally predict the trend of the infection geographically in Fortaleza and effectively supported the decision-making process of public healthcare analysts and managers from the Secretariat of Health of the State of Ceara.