Machine learning-based 4-domain framework for evaluating COVID-19 policy responses: a counterfactual analysis of 27...

European countries implemented highly diverse mitigation policies during the COVID-19 pandemic, ranging from strict nationwide lockdowns to more voluntary approaches. This study aims to understanding how timing, stringency, and comprehensiveness of policy responses influenced epidemic trajectories by applying machine learning-based counterfactual analysis.

Data from 27 European OECD countries between January 2020 and December 2022 were analysed. Daily epidemiological data, including COVID-19 cases, deaths, effective reproduction number were linked with government response indicators, demographics, vaccination, testing, and mobility data. Temporal Fusion Transformer (TFT) models were applied for multi-horizon time series forecasting by capturing nonlinear relationships between policy response indicators and COVID-19 outcome variables. Simulations were conducted under eight hypothetical response scenarios, ranging from the loosest to the strictest policy bundles. The impacts of policy responses on COVID-19 outcomes were assessed by comparing the counterfactual scenarios with the factual outcomes.

TFT models achieved excellent predictive accuracy (percentage mean absolute error <10%). The strictest responses were associated with reduced daily COVID-19 cases by over 10% in several countries, notably Hungary, Switzerland, and Turkey, while the loosest responses were associated with increased incidence by 10-20%, with the highest value observed in Poland. Associations with mortality were smaller and heterogeneous, with potentially maximum reductions of 7-10% in Belgium, Portugal, and Switzerland under strict scenarios. Over early relaxation from the restrictions might lead to outcomes as adverse as those under continuously loose policies. Feature importance analyses highlighted restrictions on mobility and gatherings, vaccination, testing, and fiscal measures as dominant drivers, alongside country-level factors such as age structure and chronic disease burden.

The TFT-based machine learning framework demonstrated favourable feasibility and interpretability, reinforcing its value in guiding policy decisions. Comprehensive, multi-domain interventions outperformed partial or short-lived restrictions. Lifting measures before achieving sufficient immunity and epidemic control posed substantial risks. Stringent policies reduced transmission, but their impact on mortality was constrained, which might be due to demographic and systemic vulnerabilities. Adaptive, data-driven strategies integrating epidemiology, policy, and structural context are essential to strengthen policy responses against future pandemics.