Evaluating predictive hybrid neural network models in spatiotemporal context: An application on Influenza outbreak predictions

Abstract

Abstract In the rigorous and varied field of infectious disease outbreak modeling, there remains a notable gap in addressing the spatiotemporal challenges present in established models. This study aimed to fill this gap by evaluating four already established hybrid neural network models designed to predict influenza outbreaks, given that influenza is a major infectious disease. These models were analyzed by employing time series data from eight different countries which is a deviation from the original articles to challenge the models with imposed spatial difficulties, in a month-on-month structure to assess their abilities to handle spatiotemporal dependencies. The models' predictions were compared using MAPE (Mean Absolute Percentage Error) and RMSE (Root Mean Squared Error) metrics, as well as graphical representations of the original values and predicted values generated by employed models. The GA-LSTM model yielded the lowest MAPE score of 62.38% followed by GA-ConvLSTM-CNN model with a MAPE score of 66.23% among all the models, and the SARIMA-LSTM model excelled in achieving the lowest average RMSE score of 66.93 as well as reporting the lowest RMSE score for three out of eight countries studied. In this case also, GA-ConvLSTM-CNN model comes in second place with an average RMSE score of 68.46. Considering these results and the ability to follow the seasonal trends of the actual values, this study suggests the SARIMA-LSTM model to be more robust to spatiotemporal challenges compared with the other models under examination.