Enhanced LSTM model with integrated attention mechanism and data augmentation for projecting COVID-19 trends in Africa

Abstract

Estimating the outcomes of epidemics promptly and precisely is crucial for decision-making and policy implementation. In this paper, we employ a long short-term memory (LSTM) method with an attention mechanism to discern the temporal correlation of COVID-19 growth. We propose a novel augmentation technique to enhance the regression model’s accuracy. A heuristic search identifies the optimal LSTM layer structure that maximizes the validation score. Initially, we trained the model on data containing confirmed cases and deaths from African countries, categorized by economic factors and GDP. The performance was better than the gated recurrent unit (GRU), LSTM, and BiLSTM methods, exhibiting a comparably low validation error. We assessed our LSTM-augmented model (LSTM-aug) using graph visualization and regression metrics on WHO COVID-19 data, demonstrating its superiority over existing methods.