Journal Article
War-driven displacement and COVID-19 in Poland: simulation study using LSTM model
Butkevych, M., Meniailov, I., Lutsiv, Y., Kraivskyi, B., Chumachenko, D.
Radioelectronic and Computer Systems, 2025:4, 25–34 (2025)
Abstract
Russia’s full-scale invasion of Ukraine led to Europe’s largest and fastest displacement since World War II. Poland received the largest inflow. Rapid movement can affect COVID-19 spread and stress testing, reporting, and vaccination systems. Aim: To evaluate whether the invasion-related displacement coincided with short-term departures in Poland’s national COVID-19 cases and deaths using an intervention-anchored counterfactual forecasting approach learned from pre-invasion trends. Object of the study: Daily COVID-19 cases and deaths in Poland. Data come from the WHO dashboard, which spans May 2020 and centers on the 24 February 2022 invasion with a 30-day post-invasion horizon. Methods: We fit a univariate stacked LSTM to pre-invasion data and forecast one step ahead for the first 30 days after 24 February 2022. The network uses LSTM(128) - LSTM(64) - Dense(25) - Dense(1) with a linear output. The timeline is split into training (before 24 January 2022), validation (24 January-23 February 2022), and testing (24 February-+30 days). Each series (cases, deaths) fits six times with different random starts. Accuracy is measured using the mean absolute percentage error (MAPE). Deviations from the counterfactual are summarized as absolute and relative effects over the 30-day window. Results: Observed daily values closely tracked the counterfactual during the first month after the invasion, with only modest, short-lived over-prediction in the middle of the window. Between the validation and test periods, the average MAPE rose from 5.94% to 14.39% for cases and from 5.90% to 14.62% for deaths, reflecting greater short-run uncertainty but no large national-level break. Conclusion: Despite exceptional migration pressure, Poland’s national COVID-19 series did not show a marked divergence from a data-driven counterfactual in the first month after 24 February 2022. Scientific novelty: To the best of our knowledge, this study provides the first Poland-focused, short-horizon, data-driven counterfactual of the invasion shock under real-world Omicron conditions. It uses a simple, transparent LSTM trained only on pre-shock national data, repeats fits to capture training variability, and quantifies departures with clear absolute and relative measures.