A Data-Driven Approach to Dengue Forecasting and Early Epidemic Detection: A Threshold Regression and Machine Learning Framework

Abstract

Dengue fever is a growing public health concern in the Philippines, with rising incidence rates and the need for more accurate early outbreak prediction tools. Traditional methods of determining epidemic thresholds often rely solely on statistical incidence rates and fail to account for dynamic factors such as climate variability and lagged effects. Building on the work of Pantolla and Gonzaga, this study introduces a system that combines threshold regression and machine learning to forecast dengue incidence and detect periods of hyperendemicity and epidemicity. The threshold regression model incorporates climate variables (rainfall, temperature, and humidity) and identifies critical thresholds (71 and 89 cases) for outbreak classification. The system also trains and compares multiple machine learning models, including SVM, Random Forest, XGBoost, ARIMA, SARIMA, RNN, GRU, and LSTM, with SVM achieving the best performance (MAPE = 33.58). This integrated approach enhances early warning capabilities and offers an easy-to-use accessible platform for local health authorities to support timely intervention and outbreak management.