Predicting dengue fever in Peninsular Malaysia using geographic information system and attention-based long short-term memory

Dengue fever is a rapidly growing public health concern, particularly in tropical and subtropical regions such as Malaysia. This thesis proposes the use of recurrent neural networks (RNNs) with integrated spatial-temporal attention for dengue fever prediction. The research empirically uses the weekly numbers of dengue cases from five Malaysian states and a set of covariates which are the climate, land use, vegetation and GIS factors. The envisaged method is based on the long short-term memory (LSTM) models, and many attention mechanisms are considered to improve the efficiency of the system. The model efficiency is assessed by the root mean squared error (RMSE) parameter. The designed approach was contrasted to five reference models which are the Random Forest, Decision Tree, Support Vector Machine (SVM), Shallow Artificial Neural Network (ANN), and Deep ANN. In the first test, LSTM models with temporal attention were better than the benchmark ones. The A-LSTM model, which is the Attention LSTM, had the best performance, with the average RMSE of 3. 663. The Stacked Attention LSTM (SA-LSTM) was second in the ranking and it had a RMSE of 3. 674. Also, both LSTM and Stacked LSTM (S-LSTM) models did almost the same, as their RMSEs were 4. 154 and 4. 133, respectively. The presented LSTM models with attention were the best among the models, hence, they outdid the benchmark models. The second experiment brought spatial attention into the picture and the model that was built using the Stacked LSTM with spatial attention (SSA-LSTM) was the one that showed the best results. It got an average RMSE 3.17, across all the lookback periods. The SSA-LSTM model in comparison with three benchmark models (SVM, DT, and ANN), the SSA-LSTM model shows a significantly lower average RMSE. Besides, the SSA-LSTM model was also tested in different states in Malaysia, and the RMSE values ranged from 2. 91 to 4. 55. The spatial attention models usually than the temporal attention ones beat the predictions of the cases of dengue. A third experiment showed that the spatial-temporal attention stacked LSTM (ST-SLSTM) model had the best performance out of all the models. It had the lowest average RMSE (2. 66), which shows that it was the most accurate in terms of the predictions. The ST-SLSTM model proved to be stable and had the capacity to be applied in different states and at different forecast horizons. Nevertheless, its complexity and computational demands were the main prohibitive factors. These research results give necessary information for the public health authorities in Malaysia and other countries that also have dengue fever issues. The LSTM models with the attention mechanisms are the basis of the proposed models that are more accurate and thus, the disease prevention and control strategies are developed.

Text FK 2024 19 - Full Text.pdf Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (4MB) Text FK 2024 19.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (4MB) Official URL or Download Paper: http://ethesis.upm.edu.my/id/eprint/18983

Actions (login required)

View Item