CNN+BiGRU-Attention Classification and TiDE-PSO Forecasting Approach for Social Media-based Predictive Analysis of Dengue

Regarding dengue hemorrhagic fever (DHF), Bandung accounts for most of Indonesian cases.One of the main obstacles in handling this disease is the delay in reporting cases to the Ministry of Health (MoH) and the discrepancy between the data reported and the actual condition in the field.This situation has an impact on the slow pace of prevention and control efforts, which in turn leads to an increase in DHF cases that lead to death.Social media is one alternative to detect the spread of the disease early.This study aims to predict the number of DHF cases by utilizing data from Facebook and X (Twitter), which are classified into three categories: infected, informative, and others.Classification is done using a hybrid method convolutional neural network (CNN), bidirectional gated recurrent unit (BiGRU), and attention layer that combines with the bidirectional encoder representations from transformers Indonesian-based (IndoBERT) model as word embedding.The classification data was then used as input to predict the number of dengue cases in Bandung.Prediction is made by combining the time series dense encoder (TiDE) method and the particle swarm optimization (PSO) algorithm.The data used included the number of DHF cases, social media data, humidity, temperature, rainfall, and wind speed from January 2014 to June 2024 weekly.The study included 155 experiments in seven scenarios to validate the classification model and eight forecasting scenarios constructed from various variables.The results show that the IndoBERT model combined with the hybrid architecture of multi-channel CNN, BiGRU, and Attention mechanism performs better than other models in multi-label classification tasks.This superiority is seen in the metrics of recall (i.e., 79%) and f1-score (i.e., 82%).Meanwhile, the TiDE-PSO model utilizing the variables of the number of previous DHF cases, humidity, temperature, rainfall, wind speed, and the social media categories "infected" became the best model for forecasting up to 24 weeks ahead.The TiDE-PSO model shows a significant improvement in performance as the length of the forecasting horizon increases.