Comparative Performance Analysis of Predictive Modelling for Pertussis Incidence in Ukraine: A Case Study of Structural Breaks in Emergency Conditions

Abstract

Pertussis remains an epidemiologically significant infection despite long-standing vaccination programs, and its dynamics in Ukraine were further distorted by the combined effects of the COVID-19 period and war-related disruption of healthcare access and surveillance. This study compared four forecasting approaches, ETS, Auto-ARIMA, Prophet, and Gradient Boosting, for modeling pertussis incidence in Ukraine under emergency conditions. Monthly national surveillance data from the Center of Public Health of the Ministry of Health of Ukraine covering 2017-2024 were analyzed. Forecast accuracy for 2024 was assessed using RMSE, MAPE, and cumulative error. The observed series demonstrated a pronounced structural break, with suppressed incidence during 2020-2022, followed by a sharp resurgence in 2024. All evaluated models reproduced the broad direction of change but failed to capture the magnitude of the outbreak, systematically underestimating both the peak and the cumulative burden. Auto-ARIMA showed the best overall performance among the tested methods, but its accuracy remained insufficient for reliable standalone forecasting in a highly unstable epidemiological environment. The findings indicate that standard time series models have limited predictive utility in the context of emergency-related structural disruption. At the same time, their comparative behavior remains informative for identifying the limits of historical extrapolation and for supporting the development of more adaptive, surveillanceaware forecasting systems for public health decision-making in crisis conditions.