Catching the Spikes: Heteroscedastic Uncertainty Quantification for Enhanced Malaria Prediction

Malaria remains a pressing global health challenge and the disease burden is further compounded by fluctuating climatic conditions. To address this problem, researchers explored adapting ConvLSTM and M-LSTM architectures to enhance malaria outbreak forecasting by utilizing key environmental indicators. %such as temperature, rainfall, and Normalized Difference Vegetation Index (NDVI) as predictors. However, these models exhibited limitations in accurately capturing sporadic spikes or irregular peaks in malaria outbreak patterns. To conquer this challenge, we utilize maximized log-likelihood of Weibull to develop a new loss function. The Weibull distribution is particularly well-suited for characterizing heavy-tailed, rare events. Our experiment on Pakistan's Malaria outbreak data spanning 2000 to 2017 showed that our new method unlocks the potential of deep learning in public health strategies but also contributes to the advancement of machine learning techniques for handling complex spatiotemporal data with irregular patterns.