Designing a Human-in-the-Loop AI System for Incident Consolidation and Severity-Aware Triage in Emergency Call Centers

Abstract

Emergency call centers experience severe cognitive and operational overload during large-scale crisis events,
driven by surges in call volume, redundant incident reports, and incomplete information. While artificial
intelligence offers opportunities to support emergency response operations, fully automated decision-making
remains inappropriate in high-stakes, time-critical contexts. This paper presents ongoing work on the design and
evaluation of a human-in-the-loop AI system that supports emergency call operators through incident
consolidation, severity-aware triage, and real-time geospatial situational awareness. The system integrates
speech-to-text transcription, natural language processing, probabilistic severity modeling, and spatial-semantic
clustering to assist operators in identifying, prioritizing, and contextualizing incoming emergency reports while
preserving human oversight. We describe the system architecture, key design decisions, and an initial technical
evaluation using a synthetic but operationally grounded emergency call dataset. Preliminary results demonstrate
promising classification performance and calibrated confidence estimates under simulated surge conditions.
Ongoing work focuses on comparative baselines, user-centered evaluation, and field-oriented validation.