Systematic Comparison of Different Compartmental Models for Predicting COVID-19 Progression

Abstract

Background/Objectives: The COVID-19 pandemic highlighted the critical need for accurate predictive models to guide public health interventions and optimize healthcare resource allocation. This study evaluates how the complexity of compartmental infectious disease models influences their forecasting accuracy and utility for pandemic resource planning. Methods: We analyzed a range of compartmental models, including simple susceptible-infected-recovered (SIR) models and more complex frameworks incorporating asymptomatic carriers and deaths. These models were calibrated and tested using real-world COVID-19 data from the United States to assess their performance in predicting symptomatic and asymptomatic infection counts, peak infection timing, and resource demands. Both adaptive models (updating parameters with real-time data) and non-adaptive models were evaluated. Results: Numerical results show that while more complex models capture detailed disease dynamics, simpler models often yield better forecast accuracy, especially during early pandemic stages or when predicting peak infection periods. Adaptive models provided the most accurate short-term forecasts but required substantial computational resources, making them less practical for long-term planning. Non-adaptive models produced stable long-term forecasts useful for strategic resource allocation, such as hospital bed and ICU planning. Conclusions: Model selection should align with the pandemic stage and decision-making horizon. Simpler models are effective for rapid early-stage interventions, adaptive models excel in short-term operational forecasting, and non-adaptive models remain valuable for long-term resource planning. These findings can inform policymakers on selecting appropriate modeling approaches to improve pandemic response effectiveness.

Keywords: COVID-19 pandemic; compartmental models; disease progression; epidemic forecasting; healthcare preparedness; prediction accuracy.

Figure A1

Bi-weekly MAE for the SIR, SEIRD, SEAIRD, and SIDARTHE models.

Figure 1

Keyword analysis with VOSviewer.

Figure 2

Comparison of actual and predicted daily COVID-19 infections in the U.S. for the SIR, SEIRD, SEAIRD, and SIDARTHE models.

Figure 3

Bi-weekly RMSE for the SIR, SEIRD, SEAIRD, and SIDARTHE models.

Figure 4

Impact of vaccination on SIDARTHE accuracy post-vaccination, i.e., post 1 April 2021.

Figure 5

Distribution of deviation from actual COVID-19 daily data in the U.S. for the dates 1 March 2020, 1 April 2021, and 1 November 2021.

Figure 6

Comparison of actual and predicted numbers of symptomatic and asymptomatic infections for the SEAIRD and SIDARTHE models, starting from 1 July 2020.

Figure 7

Predicted number of infections using the adaptive SEAIRD model (SEAIRD-A) compared to the non-adaptive model, starting from 1 November 2021.

Figure 8

Decision tree for selecting epidemic forecasting models based on data availability, epidemic stage, and forecasting goals.

Figure 9

Impact of vaccination on the total infections.

Figure 10

Impact of the hospitalization rate on the total infections.

Figure 11

Impact of the transmission rate on the total infections.