Individual Submission Summary
Share...

Direct link:

Machine Learning and Public Resource Allocation: Evidence from Emergency Medical Services in South Korea

Thursday, November 5, 1:45 to 3:15pm, Property: Boston Marriott Copley Place, Floor: 5th Floor, Room: New Hampshire

Abstract

Introduction/Background: Public resource allocation depends on information about conditions that vary across space and time. In emergency medical services (EMS), where personnel, ambulances, and station coverage are limited, resource allocation and deployment are shaped by both variations in demand and constraints in available capacity. The allocation and deployment are commonly guided by observed patterns, including historical call volumes, existing station locations, and predefined service areas that structure operational coverage. While these configurations reflect past demand and service coverage, they may not adequately capture spatial and temporal shifts in demand that affect response performance and resource utilization. This limitation becomes more pronounced under abrupt changes in demand, where reliance on past observed patterns alone may not provide a sufficient basis for effective allocation. In this context, machine learning can generate EMS demand predictions by integrating diverse local conditions, producing information not directly observable from historical patterns and enabling its use in resource allocation management.Purpose/Research Question: This study examines how machine learning-based demand forecasting can be integrated into public resource allocation management from a provider perspective. Using the case of EMS center allocation in Seoul, South Korea, the study analyzes how predicted demand can inform allocation strategies, how it can be integrated into operational decision processes, and how it affects spatial placement and the operational efficiency of resources across service areas.Methods: The analysis uses spatial data at the administrative ‘dong’ level in Seoul from 2012 to 2024. Random Forest and XGBoost machine learning models with spatial components are used to generate annual predictions of EMS call demand, which are then linked to existing service areas defined by facility locations. Using both observed patterns and model-based demand information, center-level burden is constructed and evaluated using calls per 911 officer, population, and area, with weighted Gini coefficients used to assess inequality. Allocation outcomes are compared between machine learning-based configurations and observed arrangements to assess changes in spatial placement and efficiency.Results/Findings: Among 10 centers with observed changes, 60 percent show differences exceeding an average of 2.4 kilometers between the observed relocated center locations and the optimal locations derived from predicted demand. For center-level calls per 911 officer, outcomes based on predicted demand show improved operational efficiency in most years. It decreased by 0.024 in 2020, 0.028 in 2021, and 0.029 in 2023 and 2024, indicating that model-based demand information captures operational demand patterns not reflected in observed deployment alone. Over the study period, center-level burden increases until 2022 and then declines under both approaches. However, during the COVID-19 pandemic, a model-based resource allocation method further reduced variation among centers.Conclusion/Implications: The findings suggest that machine learning-based demand predictions provide information on demand conditions that are difficult to capture from observed patterns alone, particularly during periods of abrupt change. Model-based predicted demand information can be used in the public resource allocation process, where reliance on past patterns is limited, shaping how service burden is distributed across centers.

Authors