Long Short-Term Memory-Based Multivariate Forecasting Model for Dengue: A Case Study in Sri Lanka

Dengue fever has been a significant health concern in Sri Lanka since the 1960s. The number of cases has drastically increased over the last decade. Different statistical and classical machine learning models have been proposed to forecast dengue in order to mitigate the disease from reaching its high transmission rate. More recently, the advent of neural networks has improved prognosis in an efficient manner through time forecasting of dengue data using multiple predictor variables. Understanding the strength of long short-term memory (LSTM) models in the past, this specific research has evaluated three variants of LSTM models: unidirectional LSTM, bidirectional LSTM (BiLSTM), and encoder-decoder LSTM, with the aim of predicting dengue occurrences in Sri Lanka. Weather data, including rainfall and mean temperature, were used as predictors while the efficiency of the models was assessed using RMSE. While all three models exhibited relatively good performance, the BiLSTM model significantly outperformed the other two models. This study affirms the use of LSTM for predicting dengue-like vector-borne diseases that are characterized by complex relationships with the predictors.