MULTI-LABEL CLASSIFICATION FOR SPANISH EMERGENCY TRIAGE IN SATURATED COMMUNICATION ENVIRONMENTS: A PRECISION-ORIENTED APPROACH
S. G. de-los-Cobos-Silva. Department of Electrical Engineering, Metropolitan Autonomous University – Iztapalapa, Mexico City, Mexico. E-mail: cobos@xanum.uam.mx
M. A. Gutiérrez-Andrade. Department of Electrical Engineering, Metropolitan Autonomous University – Iztapalapa, Mexico City, Mexico. E-mail: gamma@xanum.uam.mx
P. Lara-Velázquez. Department of Electrical Engineering, Metropolitan Autonomous University – Iztapalapa, Mexico City, Mexico. E-mail: plara@xanum.uam.mx
E. A. Rincón-García. Department of Electrical Engineering, Metropolitan Autonomous University – Iztapalapa, Mexico City, Mexico. E-mail: rincon@xanum.uam.mx
R. A. Mora-Gutiérrez. Department of Systems, Metropolitan Autonomous University – Azcapotzalco, Mexico City, Mexico. E-mail: mgra@correo.azc.uam.mx
E. Montes-Orozco. Departamento de Computación, Electrónica y Mecatrónica, Universidad de las Américas Puebla. San Andrés Cholula, Puebla, México E-mail: edwin.montes@udlap.mx
- Fuzzy Economic Review: Volume 30, Number 1, 2025
- DOI: 10.25102/fer.2025.01.02
Abstract
In disaster scenarios, communication networks often become congested, delaying or dropping critical emergency messages. Prioritizing life-threatening alerts is crucial to ensure limited network capacity serves the most urgent needs. This study focuses on multi-label classification of Spanish-language emergency messages, where a single text may request multiple types of assistance simultaneously. Unlike traditional methods that optimize overall accuracy, we adopt a precision-oriented approach to minimize false alarms, ensuring that alerts routed to emergency services are actionable and trustworthy.
Our methodology introduces a specialized pipeline with anti-bias text cleaning (removal of geographic entities) and a three-tier hybrid strategy to address severe class imbalance. We evaluated eight classification architectures and found that the XGBoost One-vs-Rest (OvR) model performed best, achieving a Micro F1-score of 0.725, Micro-average Precision of 0.79, and an Average Precision (AP) of 0.91 on the test set. Specifically, the model reached a precision of 0.88 for flood reports and 0.82 for medical assistance. These results indicate that the system provides a reliable automated filter for early triage, maintaining high precision even in noisy, imbalanced environments.