Forecasting Epidemic Outbreaks of Acute Respiratory Viral Infections with Early Warning Signals and Machine Learning

Forecasting critical transitions in epidemic incidence is important for epidemic surveillance. Earlier detection can support a more timely public-health response. This paper studies anomaly detection methods based on artificial intelligence and early warning signals (EWSs) in incidence time series. The goal is to predict transitions from seasonal infections to epidemic outbreaks. Two approaches are examined. The first is a classification approach that estimates whether a critical transition is near. The second is a regression approach that forecasts future infection dynamics. Several machine learning models are applied to two types of data. The models include ensemble methods (Easy Ensemble, RUSBoost, and Balanced Bagging) and deep learning architectures (Early Warning Signal Network (EWSNet), LSTM, and GRU). The first dataset contains influenza incidence with epidemic periods labeled by expert criteria. The second dataset contains unlabeled COVID-19 incidence. The results show that Easy Ensemble and EWSNet provide the best balance between precision and recall. Recurrent neural networks model the dynamics of mean values effectively, whereas variance forecasting remains more difficult. The results show that classical EWS methods and machine learning can be combined to improve epidemic forecasting and support public-health decision making.