Bridging statistics and deep learning: a causal data-fusion framework for epidemic prediction
Vasileios Ε. Papageorgiou, Γεώργιος Πετμεζάς, Aristeidis Georgakis, Nicos Maglaveras, George M. Tsaklidis · Communications in Statistics - Simulation and Computation · 2025
Several studies introduced complex deep learning (DL) models to predict the progression of the COVID-19 pandemic, aiming to effectively manage high-dimensional sequential data. In contrast, the objective of the present analysis is the investigation of the impact of integrating additional features into recurrent neural network (RNN)-type models to enhance epidemic predictive performance. Specifically, a set of 31 features is examined, encompassing environmental factors, epidemiological indicators, infection phases, and parameters estimated using particle filtering, constituting the most complete training dataset to date. These features were incorporated into six low-complexity RNN-type models, including standard, bidirectional, and time-distributed variants of gated recurrent unit (GRU) and long short-term memory (LSTM) networks. Additionally, both linear and nonlinear statistical measures, such as Granger causality and transfer entropy, were employed to assess the importance of these features in predicting new COVID-19 cases and deaths. Feature importance scores derived from these model-free measures played a critical role in guiding feature-ranking and integration into the neural networks (NNs). The findings validate the importance of incorporating additional features and quantify their contribution to forecasting accuracy, thereby enabling optimal feature selection across all six RNN-type models. Moreover, the effectiveness of transfer entropy as a feature-ranking method is underscored, with models that excluded additional features exhibiting substantial degradation in predictive performance, often reflected in RMSE increases of over 100 to 200%. Ultimately, the proposed methodology enhances the understanding of epidemic dynamics and supports more accurate and timely decision-making.