Reservoir Computing in Epidemiological Forecasting: Predicting Chicken Pox Incidence

We examine the applicability of time-series forecasting techniques to model and predict chickenpox incidence rates using publicly available epidemiological data from Rozemberczki et al. [1]. Analyzing data across both time and location is crucial in understanding disease dynamics, allowing for the identification of patterns such as temporal clustering, detection of high-incidence areas, characterization of disease spread, measurement of temporal synchrony, and forecasting future incidence rates. The primary objective of this study is to evaluate the effectiveness of neural networks in addressing this problem. Reservoir Computing, ARIMA, and various types of Recurrent Neural Networks (RNNs) have demonstrated success in tackling complex time-series issues. We assess several models based on different RNN architectures, including Long Short-Term Memory (LSTM), Bidirectional LSTM (BLSTM), Gated Recurrent Unit (GRU), Bidirectional GRU (BGRU), and compare their performance. We use a deep learning approach based on Reservoir Computing to predict chickenpox counts based on past incidence rates. We implement all the aforementioned neural network architectures for fore-casting chickenpox incidence rates and compare their prediction accuracy. Our results indicate that Reservoir Computing prediction models outperform all other models trained on the same data. Furthermore, we demonstrate that Reservoir Computing prediction models are more efficient and quicker to train and deploy in epidemiology.