Abstract
Diarrhoea (DH) disease pose significant threats to national morbidity and mortality in Vietnam, especially on children. Being a climate sensitive disease, it has strong links to various meteorological factors like rainfalls or temperatures. Hence, together with global climate changes, the risk of diarrhoea has been increasing gradually while Vietnam is already a hotspot of diarrhoea worldwide. Thus, having an effective early warning system is becoming an urgent need. However, it has not been paid enough attention with very few research works, mainly focusing on quantilizing the relationships among various climate factors and diarrhoea incidences. Exploring more sophisticated machine learning techniques is therefore an interesting work towards more efficient and effective warning systems. This paper consists of two main contributions. First, many different state-of-the-art prediction models from traditional to most recent advantaged methods, e.g., SARIMA, SARIMAX, LSTM, CNN, Xgboost, SVM, LightGBM, Catboost, LightGBM, N-HiST, BlockRNN, TCN, TFT, or Transformer, are studied for predicting DH rates for a large number of locations (55 provinces) with different climates, geographics and socio-economy factors. It provides a useful view on the overall performances of different ML models on the prediction task, which is extremely useful for other researchers when developing early-warning systems for DH in other places. Second, we introduce a novel ensemble prediction model, called dynamic weighted ensemble (DWE), for further improving the DH prediction performance. DWE is a two layer ensemble approach. The first generates different meta models based on four base component models. The second layer employs a novel approach to predict the performances of all selected meta models and uses these predicted results to dynamically combine these models in a weighted scheme to produce final results. This is totally different to traditional ensemble approaches which only rely on fixed combinations of their components. To the best of our knowledge, DWE is also the first ensemble approach for diarrhoea prediction. Extensive experiments are conducted over all 55 provinces of Vietnam to demonstrate the performance of DWE and to reveal its important characteristics.
Original language | English |
---|---|
Journal | Machine Learning |
Early online date | 20 Nov 2023 |
DOIs | |
Publication status | Early online date - 20 Nov 2023 |
Bibliographical note
Funding Information:This research is funded by Vietnam National University HoChiMinh City (VNU-HCM) under Grant Number DS2022-26-03.
Publisher Copyright:
© 2023, The Author(s), under exclusive licence to Springer Science+Business Media LLC, part of Springer Nature.
Keywords
- CNN
- Deep learning
- Diarrhoe incident prediction
- Dynamic weighted ensemble
- LSTM
- LSTM-ATT
ASJC Scopus subject areas
- Software
- Artificial Intelligence