Deep Learning Architectures Framework for Emerging Outbreak Forecasting of Mpox: A Bagged Ensemble Scheme to Model Accurate Prediction Intervals

Abstract Background The Mpox virus is a disease of rare occurrence from the same family as the variola virus, which is hardly ever fatal, and its symptoms are like the ones of smallpox. As the outbreak was emerging in Brazil, the frail post-pandemic public health system and stigmatization yielded substandard data collection of confirmed cases. As of the beginning of 2023, the established case trend is short and has noisy patterns that challenge most existing forecasting methods. Methods To extend the modeling choices for emerging outbreaks with volatile and short-term confirmed case data, we evaluate the performance of multiple deep learning architectures, including Convolutional Neural Network (CNN), unidirectional Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM). We further developed a bagged ensemble methodology (i.e., BaLSTM) with the best-identified method for predicting the Mpox emerging outbreak in Brazil. Our evaluation dataset consists of Brazil’s weekly Mpox cases from July 2022 to January 2023, contrasting the performance for the coming six weeks. Additionally, accepted machine-learning models were built to predict weekly confirmed cases to compare the relative performance of our implementation. We evaluated the performance of our deep learning architectures with exponential smoothing (ETS), ARIMA, Support Vector Machine (SVM), K-nearest neighbors (KNN), and Neural Networks Autoregression (NNAR). Results Based on the results, the BaLSTM approach achieves an accuracy of 80.83% when considering its associated prediction intervals, translating into a 45% improvement in measuring forecast errors for the Mpox Brazil cases when evaluated with the remaining assessed methods. The proposed model can capture trends and patterns in the time series while considering the intrinsic attributes of the sample. Conclusions Implementation such as those described in this research will become increasingly crucial in predicting emerging epidemics with a small sample size and an apparent uncertain behavior.