Predictive Modelling of Incident Risk to Pre-empt Risk in Highway Operations: A Machine Learning Approach

Bortey, Loretta (2025) Predictive Modelling of Incident Risk to Pre-empt Risk in Highway Operations: A Machine Learning Approach. Doctoral thesis, Birmingham City University.

Preview

Text Loretta Bortey PhD Thesis_Final Version_Final Award May 2025.pdf - Accepted Version Download (6MB)

Abstract

Highway traffic officers (HTOs) operate in complex and hazardous environments, yet transportation safety research has predominantly focused on drivers, pedestrians, roads, and vehicles, with limited attention to HTOs' safety. In the UK, National Highways currently employs traditional statistical methods to mitigate safety risks post-incident, a reactive approach that does little for risk prevention. This thesis proposes a proactive approach by developing a machine learning (ML) prediction model to forecast incidents such as injuries, incursions and environmental hazards, assess risk levels, and predict the body parts likely to be affected in injurious events. The aim is to provide highway safety authorities with predictive insights for timely interventions and enhanced risk management. Despite the growing application of ML in safety risk prediction, there is limited evidence on the reliability of variables used as indicators of safety performance. To address this gap, this study develops a conceptual framework for selecting optimal safety indicators (SIs) and formulating input variables that enhance ML-based risk prediction. A three-stage, multiphase mixed-methods research design was employed: i) developing the conceptual framework; ii) constructing the proof-of-concept ML model; and iii) validating the model’s performance. The conceptual framework was established through a systematic literature review using PRISMA-based bibliometric search, scientometric and cluster analysis to identify significant SIs and grounded theory analysis was used for synthesis. The ML model development phase applied supervised learning algorithms, including Support Vector Machine (SVM), Random Forest (RF), Naïve Bayes (NB), Deep Neural Networks (DNN), Ensemble Learning (EL), and Recurrent Neural Networks (RNN). The models were trained using secondary data from a highway incident database and three data balancing techniques were tested to address the class imbalance. Model validation employed a stratified k-fold cross-validation approach, evaluated based on AUROC, precision, recall, and accuracy.

The study identifies key considerations for selecting SIs, emphasizing the integration of leading and lagging indicators to enhance system adaptability and resilience. A novel conceptual framework is presented that guides the selection of robust indicators for ML-based risk modelling. Empirical findings indicate that the SVM model with a polynomial kernel, combined with the SMOTE algorithm, outperforms other models in predicting incident types, risk levels, and affected body parts, whereas Random Under-sampling (RU) was the least effective. Critical factors influencing highway incidents, including weather conditions, visibility, age range and location, were identified and analysed.

This research makes several novel contributions: i) a novel conceptual framework integrating resilient SIs for predictive modelling; ii) a systematic approach to combining leading and lagging indicators for enhanced safety performance; and iii) the first study to use an incident database dedicated to HTOs for predictive risk modelling. The developed ML model provides actionable insights for safety officers, enabling proactive risk mitigation through targeted training and preparedness strategies for HTOs. This ultimately improves workplace safety in highway operations.

Item Type:	Thesis (Doctoral)
Dates:	Date Event 14 May 2025 Accepted
Uncontrolled Keywords:	AI, Machine learning, predictive modelling, safety risk, highway
Subjects:	CAH10 - engineering and technology > CAH10-01 - engineering > CAH10-01-07 - civil engineering CAH11 - computing > CAH11-01 - computing > CAH11-01-05 - artificial intelligence
Divisions:	Architecture, Built Environment, Computing and Engineering > Architecture and Built Environment > Built Environment Doctoral Research College > Doctoral Theses Collection
Depositing User:	Louise Muldowney
Date Deposited:	18 Aug 2025 09:53
Last Modified:	18 Aug 2025 09:53
URI:	https://www.open-access.bcu.ac.uk/id/eprint/16597

Actions (login required)

View Item