Inverse problem for parameters identification in a modified SIRD epidemic model using ensemble neural networks.

Area of Science:

• Epidemiology
• Computational Biology
• Machine Learning

Background:

• The Susceptible-Infectious-Recovered-Deceased (SIRD) model is a key tool for understanding infectious disease dynamics.
• Traditional SIRD models often assume constant parameters, which is unrealistic given real-world factors like policy changes and viral variants.
• Accurate short-term forecasting is crucial for effective public health interventions.

Purpose of the Study:

• To develop a dynamic parameter identification methodology for the SIRD model suitable for short-term predictions.
• To incorporate a parameter accounting for the discrepancy between reported and actual infected cases.
• To apply and validate the methodology for COVID-19 forecasting in Romania and other European countries.

Main Methods:

• A novel parameter identification methodology for the SIRD model was developed.
• An ensemble of neural networks was trained on a synthetic dataset generated by solving the SIRD model with randomized parameters.
• The methodology utilizes the past 7 days of data for parameter estimation and subsequent predictions.

Main Results:

• The proposed method successfully estimated SIRD model parameters using real COVID-19 data from Romania.
• Predictions for the number of deaths were generated for periods ranging from 10 to 45 days.
• The methodology demonstrated similar efficacy when applied to data from Hungary, the Czech Republic, and Poland.

Conclusions:

• The developed methodology provides a robust approach for short-term infectious disease forecasting.
• The parameter identification technique is adaptable for various compartmental models and infectious diseases.
• A supporting theorem validates the ability to recover model parameters from reported data.