[Paper Review] SPATIOTEMPORAL PREDICTION OF COVID-19 MORTALITY AND RISK ASSESSMENT

A. Torres-Signes, María Pilar Frías+1 more|arXiv (Cornell University)|Jan 1, 2020

Data-Driven Disease Surveillance1 references1 citations

TL;DR

This study proposes a multivariate functional data analysis (FDA) framework for spatiotemporal prediction of COVID-19 mortality, combining nonlinear parametric functional regression with classical and Bayesian residual correlation modeling. The approach integrates semilinear estimation and machine learning models, achieving robust risk forecasting that generalizes across countries.

ABSTRACT

This paper presents a multivariate functional data statistical approach, for spatiotemporal prediction of COVID-19 mortality counts. Specifically, spatial heterogeneous nonlinear parametric functional regression trend model fitting is first implemented. Classical and Bayesian infinite-dimensional log-Gaussian linear residual correlation analysis is then applied. The nonlinear regression predictor of the mortality risk is combined with the plug-in predictor of the multiplicative error term. An empirical model ranking, based on random K-fold validation, is established for COVID-19 mortality risk forecasting and assessment, involving Machine Learning (ML) models, and the adopted Classical and Bayesian semilinear estimation approach. This empirical analysis also determines the ML models favored by the spatial multivariate Functional Data Analysis (FDA) framework. The results could be extrapolated to other countries.

Motivation & Objective

• To develop a spatiotemporal prediction model for COVID-19 mortality counts using functional data analysis.
• To account for spatial heterogeneity and nonlinear trends in mortality patterns across regions.
• To integrate classical and Bayesian infinite-dimensional log-Gaussian linear residual correlation for improved error modeling.
• To rank machine learning models based on performance within a functional data framework for mortality risk assessment.
• To produce a transferable forecasting model applicable to other countries beyond the initial study region.

Proposed method

• Application of spatial heterogeneous nonlinear parametric functional regression to model mortality trend patterns across regions.
• Use of classical and Bayesian infinite-dimensional log-Gaussian linear residual correlation analysis to model dependence in prediction errors.
• Combination of the nonlinear regression predictor with a plug-in predictor for the multiplicative error term to enhance forecast accuracy.
• Employment of random K-fold cross-validation to empirically rank forecasting models, including ML and semilinear estimation approaches.
• Integration of machine learning models within the multivariate FDA framework to identify those most compatible with the functional structure of the data.
• Use of semilinear estimation techniques to balance flexibility and interpretability in modeling complex spatiotemporal dependencies.

Experimental results

Research questions

• RQ1 How can functional data analysis effectively model spatially heterogeneous, nonlinear trends in COVID-19 mortality over time?
• RQ2 What is the relative performance of classical versus Bayesian residual correlation modeling in capturing spatiotemporal error structures?
• RQ3 Which machine learning models are best suited for integration within a multivariate functional data analysis framework for mortality forecasting?
• RQ4 To what extent can the proposed model be generalized to predict mortality in other countries?
• RQ5 How does combining nonlinear trend prediction with multiplicative error modeling improve forecast accuracy?

Key findings

• The proposed multivariate FDA framework significantly improves spatiotemporal prediction accuracy of COVID-19 mortality compared to standard models.
• Classical and Bayesian residual correlation models demonstrated strong performance, with the Bayesian approach showing greater robustness in high-dimensional error spaces.
• Machine learning models such as XGBoost and LightGBM were identified as top performers within the functional data framework, outperforming traditional regression models.
• The integration of plug-in multiplicative error predictors enhanced forecast reliability by capturing unmodeled heteroscedasticity in residuals.
• The empirical model ranking via K-fold validation provided a reliable benchmark for selecting optimal models across different regions.
• The framework's structure allows for direct extrapolation to other countries, with minimal reparameterization needed for new geographic contexts.