Assessing the Efficacy of Pre-trained Large Language Model for Intersection Crash Severity Classification

Traffic safety analysis plays a critical role in preventing road crashes by identifying and addressing risk factors before accidents occur. Crash severity prediction is a critical aspect of traffic safety research, yet traditional machine learning (ML) models often demand labor-intensive preprocessing and produce results that are often complex to interpret for non-technical stakeholders. This study investigates the potential of large language models (LLMs), specifically GPT-4o, for crash severity classification using a few-shot prompting approach that bypasses resource-intensive fine-tuning. Structured crash data from the North Dakota Department of Transportation (2013–2023) were transformed into narrative-style prompts through a tabular-to-text framework, enabling GPT-4o to process real-world crash reports in natural language. The model’s performance was benchmarked against five classical ML algorithms Logistic Regression, Random Forest, Decision Tree, Naive Bayes, and CatBoost using accuracy, precision, recall, and F1 score. Results show that GPT-4o achieved the highest performance across all metrics (accuracy: 0.50, precision: 0.68, recall: 0.50, F1 score: 0.50), outperforming traditional ML baselines under minimal preprocessing. Compared to a prior study using a fine-tuned LLaMA-2 70B model, GPT-4o demonstrated a relative improvement of 11.1% in accuracy and F1, 4.2% in recall, and a substantial 51.1% in precision. These findings highlight the practicality of LLMs for crash severity analysis in low-resource settings, offering scalable, interpretable, and plug-and-play alternatives to conventional ML methods. With simplified deployment and interpretation, this approach provides transportation agencies a practical pathway to adopt AI for safety analysis, as the use of natural language processing lowers technical barriers significantly.