Severity-aware optimization of UAV-based emergency medical services with AI-driven prioritization

Abstract

Rapid emergency medical response following disasters is often hindered by damaged infrastructure, limited situational awareness, and the difficulty of rapidly assessing and prioritizing victims using conventional emergency medical service (EMS) systems. Although uncrewed aerial vehicles (UAVs) have shown promise for aerial reconnaissance and disaster monitoring, existing UAV-assisted emergency-response frameworks typically focus either on victim detection or on logistics-oriented resource allocation in isolation, with limited integration between aerial perception and downstream dispatch decision making. Consequently, current systems do not adequately support severity-aware UAV-assisted EMS allocation in which dispatch decisions are informed by the inferred condition or urgency of observed victims. To address this problem, this thesis proposes an integrated UAV-assisted emergency medical response framework that links aerial victim detection, visual criticality estimation, and optimization-based UAV dispatch within a unified perception-to-decision pipeline. UAV-acquired disaster imagery is first processed using a YOLOv8-based human detection model, a deep learning–based real-time object detection algorithm, to localize affected individuals. Detected victims are then analyzed using a binary criticality classifier trained on aerial disaster imagery from the C2A dataset, augmented with posture-based criticality annotations to distinguish higher-risk victims from less urgent cases. These outputs are combined within a triage-inspired scoring framework to generate severity and priority estimates for spatial demand regions. The resulting perception-derived severity and priority information is incorporated into a tailored mixed integer linear programming (MILP) model for UAV-enabled EMS dispatch and facility-allocation optimization that jointly considers travel time, operational cost, severity coverage, and priority coverage. Unlike conventional cost-focused UAV-assisted EMS baseline, which assumes homogeneous demand, the proposed model explicitly incorporates perception-derived triage information into dispatch decisions. Experimental evaluation demonstrates that incorporating perception-derived severity and priority information enables the proposed framework to allocate UAV resources in a manner more aligned with victim criticality than conventional cost-focused dispatch strategies. These results demonstrate the feasibility of integrating aerial perception with optimization-based dispatch to support severity-aware UAV-assisted EMS planning and provide a foundation for future perception-driven emergency-response systems.