Comparative evaluation of machine learning and LLM-based approaches for construction accident prediction

High injury and fatality rates in the construction industry continue to result in significant human and economic losses. To improve construction safety, construction accident prediction has become a critical component of safety management by leveraging machine learning (ML) and deep learning (DL) approaches based on historical datasets. However, they are limited to capturing contextual causality within datasets and performing well in new contexts. A promising alternative is to use large language models (LLMs) that enable data-based causal reasoning and capture contextual pre-incident factors. Unfortunately, the feasibility of using LLMs to predict construction accidents has not been sufficiently investigated, leaving several research gaps (e.g. prompting strategies, performance relative to ML methods, and validation) unresolved. Therefore, this study aims to demonstrate feasibility by developing two prediction models (ML and LLM), comparing their performance, and validating them. Fourteen different ML algorithms were used, and a separate LLM classifier was developed using Chain-of-Thought prompting combined with the Issue, Rule, Application, and Conclusion (IRAC) approach, which is effective in shaping logical reasoning processes. Results showed that XGBoost and LLM achieved accuracies of 97.9% and 94.5%, and recalls of 97.5% and 98.9%, respectively. Also, validation with 1,206 hold-out datasets confirmed temporal robustness for both approaches.