This paper addresses the critical issue of predicting traffic accident durations on expressways, which is essential for traffic management and emergency response. The authors propose a heterogeneous deep learning architecture that employs multi-modal features to enhance prediction accuracy. The study involves six unique data modes, including structured data and text data, and uses a hybrid deep learning approach to build classification models with reduced prediction error. The influence of multi-modal data on accident duration prediction performances is rigorously analyzed using various deep learning models. The proposed method is evaluated using survey data from an expressway monitoring system in Shaanxi Province, China. The results show that the Word2Vec-BiGRU-CNN model achieves the highest F1-score of 0.3648, demonstrating the effectiveness of using text features for traffic accident duration prediction. The study confirms that structured features extracted from text data significantly improve the prediction effects of deep learning algorithms, while these features may be detrimental to conventional machine learning algorithms. The research highlights the importance of processing and extracting text features in traffic accident duration prediction and suggests that future work should explore additional multimodal data fusion methods to further enhance prediction accuracy.This paper addresses the critical issue of predicting traffic accident durations on expressways, which is essential for traffic management and emergency response. The authors propose a heterogeneous deep learning architecture that employs multi-modal features to enhance prediction accuracy. The study involves six unique data modes, including structured data and text data, and uses a hybrid deep learning approach to build classification models with reduced prediction error. The influence of multi-modal data on accident duration prediction performances is rigorously analyzed using various deep learning models. The proposed method is evaluated using survey data from an expressway monitoring system in Shaanxi Province, China. The results show that the Word2Vec-BiGRU-CNN model achieves the highest F1-score of 0.3648, demonstrating the effectiveness of using text features for traffic accident duration prediction. The study confirms that structured features extracted from text data significantly improve the prediction effects of deep learning algorithms, while these features may be detrimental to conventional machine learning algorithms. The research highlights the importance of processing and extracting text features in traffic accident duration prediction and suggests that future work should explore additional multimodal data fusion methods to further enhance prediction accuracy.