Dynamic feature selection improves influenza forecasting accuracy and generalization across countries

INTRODUCTION: The differentiation in epidemic patterns and multiple influencing factors pose significant challenges to influenza forecasting, highlighting the need for novel methods to improve predictive accuracy and cross-regional generalizability. OBJECTIVES: This study aims to develop an adaptive feature selection model named AdaFluDR to address the time-varying nature of influenza transmission drivers across different periods and regions. METHODS: AdaFluDR integrates the SpaceTime and Crossformer models and utilizes a correlation-driven mechanism. This mechanism constructs a comprehensive score by integrating the temporal, frequency, and time domain information of features, and dynamically adjusts feature processing pathways based on this score. Subsequently, a multilayer perceptron (MLP) will be employed to model the nonlinear mapping relationship between the integrated features and the target variable for prediction generation. RESULTS: The AdaFluDR model outperforms traditional methods and other machine learning approaches, demonstrating robust predictive performance across multiple forecasting horizons (1-4 weeks), and strong generalization ability across the United States, Canada, and Portugal. CONCLUSION: Our study provides a novel and practical framework for forecasting influenza activity with reliable accuracy and cross-national applicability, providing a valuable tool for improving global epidemic preparedness and response strategies.