A Sparse Representation of Social Media, Internet Query, and Surveillance Data to Forecast Dengue Case Number using Hybrid DecompositionBidirectional LSTM

Abstract

Dengue fever is an endemic disease that occurs throughout the year. Forecasting cases of dengue fever based on actual data is needed for monitoring and taking action. Recently, developing countries have faced problems related to the dengue fever surveillance system caused by the data delay factor. On the other hand, availability and access to health-related information on the internet have changed people's behaviors and habits. However, the effect of internet data usage has not been widely studied, especially in areas with different levels of internet penetration. This study examines the impact of dengue fever case reported data, Google Trends, Twitter, and climate data in areas with many cases and varying levels of internet penetration to forecast dengue fever cases. Split time-series cross-validation (STSCV) and blocked time-series cross-validation (BTSCV) are used to obtain various training and testing results. The hybrid Decomposition-Bidirectional Long Short-Term Memory(D-BiLSTM) method is proposed. D-BiLSTM applied to eight different scenarios across multiple level areas. According to the results of the experiments, the D-BiLSTM model with STSCV outperforms the BTSCV. In the high internet penetration area, the average error is 9,517, while in the low internet penetration area, it is 5,188. In areas with high internet penetration, adding the variables Google Trends and Twitter does not significantly reduce the error forecasting. However, in the low penetration area, the inclusion of Google Trends and Twitter significantly decreases errors. In general, the D-BiLSTM model performed well. Then, when compared with other approaches, the D-BiLSTM model as a whole can reduce the average RMSE and the average MAE of the comparison model by 94,120 and 45,132, respectively, in areas of high internet penetration with the best SMAPE model of 0.310. In the low internet penetration area, the average decline in RMSE and MAE was 54,390 and 19,362, with the best SMAPE model performance of 0.183.

Combination of BERT and Hybrid CNN-LSTM Models for Indonesia Dengue Tweets Classification

“…Tweeter is said to be the most important corpus source for text mining in the case of dengue illness [9]. Tweet data has also been used to predict dengue cases at the local and national levels in Brazil [34] and in various locations of Indonesia [18,19]. Tweet data is paired with case count data to forecast dengue fever cases during the coming time.…”

“…Tweet data is paired with case count data to forecast dengue fever cases during the coming time. The use of Twitter data in [19,21] concentrates on the development of an intelligent indicator system from Twitter that uses sentiment analysis to detect dengue fever incidents in India. Additionally, [17] utilized polarity sentiment analysis to acquire data and extract pertinent information regarding dengue fever.…”

“…Tweet data is also processed to provide data sets that may be useful in the domains of epidemiology and infodemiology [11]. In the case of infectious diseases, Twitter data is commonly utilized to conduct sentiment analysis, detection, or prediction of disease transmission or trends on the topics of malaria [12], Zika [13], COVID-19 [14][15][16][17], and dengue fever [18][19][20][21][22][23].…”

“…There has been some research using Twitter data for dengue fever cases in recent years, but to the best of the author's knowledge, Twitter data on dengue fever cases are not widely used for classification purposes, particularly text-based classification in Indonesian. In fact, the results of the classification of dengue-related news tweets can be used for nowcasting functionality of current disease levels [20], forecasting [19] and early detection of an outbreak [21], as well as identifying the top users who are actively publishing relevant content about the topic [26,27]. Twitter has been used in [22,23] to track dengue fever cases in Indonesia.…”

Internet search data with spatiotemporal analysis in infectious disease surveillance: Challenges and perspectives

“…Traditional data analysis approaches often generate correlations with biases under the assumption that the independent variables have no autocorrelation at both spatial and temporal scales ( 37 ). However, such autocorrelation is very common in the real world.…”