Enhancing malaria prediction in Amazonas, Brazil: A comparative analysis of deep learning models based on clustering

Abstract

Abstract Backgound: Malaria is a disease transmitted by the female Anopheles mosquito and can be treatable. However, in 2019 there were over 247 million instances of malaria reported, along with 619,000 fatalities. More than 42 million Brazilians are at danger of developing malaria, with 99 percent of cases occurring in or around the Amazon forest (WHO, 2020). In Brazil, malaria is still a significant public health issue, despite the decline in occurrences and deaths. Methods: Accurate spatiotemporal forecasting of malaria transmission could help to better allocate resources to help fight the disease. In order to estimate malaria cases in the Amazonas state, we compare and evaluate deep learning, machine learning, and statistical models in this work. We used k-means clustering to group municipalities using a Brazilian dataset of around 6 million records (January 2003 to December 2019) to test if the performance of the models may be improved when grouping municipalities with statistically similar incidence rates of malaria. Results: The results indicate that the ARIMA model achieved better performance, but, the other models obtained similar values. The division of municipalities by clusters reinforced the application of models for municipalities with similar statistical values. Conclusions: As for geographic distribution, the results are similar. However, evaluating municipalities by grouping K-means, based on statistical values, is an alternative for decision-makers, as it will be necessary to monitor only 5 large groups instead of 9 health regions. The behavior of the models comparing the dataset with all types of malaria and Malaria Vivax was similar since the amount of data available for the models to perform the learning is quite close.