Comparative study on influenza time series prediction models in a megacity from 2010 to 2019: Based on SARIMA and deep learning hybrid prediction model

Abstract Background: It is very important to establish more accurate influenza prediction models in megacities. The purpose of this study was to compare the different time series prediction models for influenza from 2010 to 2019 in Beijing, China. Methods: We took the influenza-like illness rate (ILI%), the influenza positive rate and the product of ILI% and influenza positive rate as dependent variables respectively. Subsequently, and model performances of summer point, peak bottom point and peak rising point were analyzed. After selecting the best prediction point, we compared the model performances of different parameters at that point using the SARIMA model. Then, the best model selected by SARIMA was compared with the hybrid LSTM model. Results: Between the 26th week of 2010 and the 25th week of 2019, there were 6,753,116(1.24%) ILI patients, 15,883(16.75%) of which were positive for influenza.The trends and the peak times of ILI%, the influenza positive rate and the product of ILI% and influenza positive rate were roughly the same. The SARIMA model of the peak rising point was better than those of the summer point and peak bottom point. The hybrid LSTM model performed better than the selected best SARIMA model in terms of ILI%, influenza positive rate and the product of ILI% and influenza positive rate. Also, the hybrid LSTM model could maintain a good prediction effect from the 1st to the 26th week.. On the contrary, the prediction effect of the SARIMA model decreased significantly with the extension of the prediction period. Conclusions: Our results suggested that the prediction effect of the hybrid LSTM model was better than the SARIMA model, in terms of ILI%, influenza positive rate and the product of ILI% and influenza positive rate. SARIMA was more suitable for short-term prediction, while the hybrid LSTM model showed obvious advantages in long-term prediction. Our research could help to improve the prediction and early warning of influenza and other respiratory infectious diseases.