Hybrid modeling approaches for predicting COVID-19 mortality: A comparative study across USA, France, and India

The precise forecasting of COVID-19 mortality is essential for prompt public health measures and resource distribution. This paper proposes and assesses three Gaussian Process-based hybrid models—GP-RBM (Random Boosting Machine), GP-LSTM (Long Short-Term Memory), and GP-CNN (Convolutional Neural Network)—for forecasting COVID-19 mortality cases in the USA, France, and India. The models combine the adaptability of Gaussian Processes with the deep learning functionalities of RBM, LSTM, and CNN to improve prediction precision and uncertainty assessment. We utilize various evaluation metrics to determine the optimal hybrid model: Symmetric Mean Absolute Percentage Error (sMAPE) for accuracy, Continuous Ranked Probability Score (CRPS) for probabilistic forecasts, and Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) for model selection. Furthermore, we examine the Peak Timing Error and Peak Magnitude Error to ascertain the models' proficiency in capturing essential epidemiological patterns. The Diebold-Mariano Test is used to statistically compare the predicting performance of models. The findings indicate notable disparities in prediction accuracy, with each model exhibiting both advantages and drawbacks across different epidemiological contexts. The GP-LSTM and GP-CNN hybrid models surpassed the GP-RBM model in accuracy and efficiency, especially with the datasets from France and India. This comparative research underscores the need of choosing suitable hybrid models customized to the unique dynamics of regional COVID-19 data. The results provide significant insights into the efficacy of sophisticated machine learning methods in epidemic forecasting, presenting pragmatic recommendations for future pandemic readiness. • This paper utilized three Gaussian Process-based hybrid models GP-RBM, GP-LSTM, and GP-CNN for forecasting COVID-19 mortality. • The deep learning functionalities of RBM, LSTM, and CNN to improve prediction precision. • We utilize sMAPE for accuracy, CRPS for probabilistic forecasts, AIC and BIC for model selection. • The findings indicate the GP-LSTM and GP-CNN models surpassed the GP-RBM model in accuracy and efficiency.