A Model-Free Geotemporal Clustering Approach for Pandemic Projection and Medical Resource Planning: Insights from COVID-19

Abstract

Problem definition: We propose a data-driven approach for pandemic projection that requires minimal data and operates without disease-specific assumptions. This approach is integrated into an optimization framework for medical personnel planning during pandemics. Methodology/results: Our algorithm employs geotemporal clustering to predict pandemic trajectories. Each (location, date) pair is characterized by time-series features and grouped via k-means clustering. Future trends are predicted by linking current observations to similar historical patterns within clusters. We further leverage these predictions to develop an optimization method for medical personnel planning. In forecasting COVID-19 cases, deaths, and hospitalizations across the United States, our algorithm outperforms several established but more complex models. Furthermore, incorporating our algorithm enhances the performance of existing models. For medical personnel planning, our optimization approach achieves substantial cost reductions compared to the sample average approximation (SAA) method. Managerial implications: Our data-light, model-free approach provides reliable real-time pandemic projections while requiring minimal historical data. This enables healthcare institutions to make informed staffing decisions even when data is limited or rapidly changing. The method is particularly valuable during the early stages of public health crises when detailed epidemiological information may be unavailable.

Note: Funding: None to declare

Declaration of Interest: None to declare

Keywords: pandemic projection, medical personnel planning, clustering, sample average approximation, COVID-19