Logistic regression vs machine learning to predict evacuation decisions in fire alarm situations - research article

Aug 3, 2026

Upload by Maeve

This study evaluates logistic regression and machine learning models for predicting evacuation decisions in fire alarm situations and compares their predictive performance. Seven algorithms are trained and calibrated using web-based experimental data from 1,807 participants, and model quality is assessed with AUC, accuracy, recall, specificity, precision, and F1-score. Results show logistic regression achieves the highest AUC (0.831), while extreme gradient boosting leads accuracy (0.780), specificity (0.810), and precision (0.820). K-nearest neighbours yields the highest recall (0.859), and the artificial neural network attains the best F1-score (0.785).

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Introduction

Decision-making in emergencies and the stay/go choice

Warning systems and delayed response

Influencing factors in evacuation studies

Decision-making research on wildfires

Methods and model evaluation

Machine learning algorithms and training data

Performance metrics and comparisons

Results and key influential factors

Safety Science 174 (2024) 106485 Contents lists available at ScienceDirect Safety Science journal [homepage: www.elsevier.com/locate/safety](homepage: www.elsevier.com/locate/safety) | Logistic regression vs machine learning to predict evacuation decisions in fire alarm situations
*
Adriana Balboa , Arturo Cuesta , Javier Gonz´alez-Villa , Gemma Ortiz , Daniel Alvear
Universidad de Cantabria, Los Castros s/n, Santander 39005, Spain | | | | --- | --- | --- | | A R T I C L E I N F O | | A B S T R A C T | | Keywords: Evacuation Decision-making Logistic regression
Machine learning Fire alarm | | In this study we assessed logistic regression and machine learning models to explore their performance in predicting evacuation decisions and to provide readers with insights into the accuracy of these methods. We tested seven machine learning algorithms, including classification and regression tree, Naïve Bayes, K-nearest neighbours, support vector machine, random forest, extreme gradient boosting, and artificial neural network. We used data collected from 1,807 participants through web-based experiments to train and calibrate these models. The performance of each model was evaluated by area under the curve, accuracy, recall, specificity, precision, and F1-score. The results indicate that logistic regression had the highest area under the curve value (0.831), whereas extreme gradient boosting outperformed other machine learning models in terms of accuracy (0.780), specificity (0.810) and precision (0.820). K-nearest neighbours model had the greater recall (0.859) and artificial neural network the highest F1-score (0.785). The models identified that being with a close person was the most influential factor in the response to a fire alarm. | 1. Introduction Understanding decision making of people during an emergency is a main concern in safety science (Kuligowski, 2009; Kuligowski, 2011; Proulx, 2001; Wood, 1972). The “stay” or “go” decision of individuals is a crucial aspect of the pre-evacuation phase, especially when traditional warning signals such as alarm bells, horns, or sirens are used. These warning systems usually fail to create an immediate response (Proulx, 2001; Wood, 1972) as they only inform occupants about the potential danger without providing any further information of the emergency. Consequently, people may either ignore the warning and carry on with their activities or seek additional information (Proulx, 2001; Wood, 1972), resulting in a delayed response that can increase the risk to life safety (Fritz and Marks, 1954; Kuligowski, 2009). To address this, researchers have focused on improving communication systems for warning individuals (D’Orazio et al., 2016; Ferraro and Settino, 2019; Kuligowski, 2011; Lin et al., 2023), evaluating the decision-making during pre-movement phase to support fire safety assessment (Liu and Lo, 2011; Lovreglio et al., 2015; Viswanathan and Lees, 2014) and analysing wayfinding behaviour (Lin et al., 2019; Vilaret al., 2018). Enhancing our knowledge of the factors that influence human behaviour during emergencies is critical (Santos and Aguirre, 2004), and therefore, the analysis of these factors should be considered in evacuation studies and plans to avoid biased analysis (Song and Lovreglio, 2021). Previous research has already begun to analyse the impact offactors on human behaviour during building fires. Studies have examined the physical context, including the environment (Kinateder et al., 2018; Richardson et al., 2018; Cubukcu, 2003), and cues (Yamada and Akizuki, 2016; Fu et al., 2018; Saunders, 2001). Social influences such asthe presence or absence of others (Fu et al., 2018; Song and Lovreglio, 2021; Lovreglio et al., 2015) have also been explored. Demographics such as age, education or gender have also been studied (Jeon et al., 2014; Song and Lovreglio, 2021; Saunders, 2001). Furthermore, a significant body of literature has investigated decision

FAQs

What models are compared for predicting evacuation decisions in this study?

The study compares logistic regression with seven machine learning algorithms, including decision tree methods, Naïve Bayes, k-nearest neighbours, support vector machine, random forest, extreme gradient boosting, and an artificial neural network.

How is model performance evaluated?

Performance is evaluated using area under the curve (AUC) and classification metrics including accuracy, recall, specificity, precision, and F1-score.

Which factor is identified as most influential for responses to a fire alarm?

Being with a close person is identified as the most influential factor affecting responses to a fire alarm.

Aug 3, 2026

Upload by Maeve