This research article proposes a heterogeneous deep learning architecture that integrates multi-modal features to improve the accuracy of traffic accident duration predictions on expressways. The study addresses the limitations of existing methods that primarily rely on single-mode data, and investigates the impact of multi-mode data on prediction performance. The proposed method combines structured data and text data to extract new structured features, which are then used to build classification models with reduced prediction error. A hybrid deep learning approach is applied to enhance the prediction accuracy, and the influence of multi-mode data on accident duration prediction is analyzed using various deep learning models. The study evaluates the proposed method using survey data collected from an expressway monitoring system in Shaanxi Province, China. The results show that the Word2Vec-BiGRU-CNN model achieves a high F1-score of 0.3648 for traffic accident duration prediction. This study confirms that the newly established structured features extracted from text data significantly enhance the prediction effects of deep learning algorithms. However, these new features are detrimental to the prediction effects of conventional machine learning algorithms. The results demonstrate that text feature processing and extraction is a complex issue in traffic accident duration prediction. The proposed method uses a BiGRU-CNN architecture to process multi-modal data, including structured data and text data. The model is trained using six defined data modes, which are derived from the dataset. The results show that the BiGRU-CNN model provides the best prediction performance with the highest mean-F1 value when working with text data. The study concludes that the integration of multi-modal data and deep learning models can significantly improve the accuracy of traffic accident duration predictions. The results also highlight the importance of text vectorization in enhancing the performance of deep learning algorithms. The study suggests that future research should explore more advanced data fusion methods to further improve prediction accuracy.This research article proposes a heterogeneous deep learning architecture that integrates multi-modal features to improve the accuracy of traffic accident duration predictions on expressways. The study addresses the limitations of existing methods that primarily rely on single-mode data, and investigates the impact of multi-mode data on prediction performance. The proposed method combines structured data and text data to extract new structured features, which are then used to build classification models with reduced prediction error. A hybrid deep learning approach is applied to enhance the prediction accuracy, and the influence of multi-mode data on accident duration prediction is analyzed using various deep learning models. The study evaluates the proposed method using survey data collected from an expressway monitoring system in Shaanxi Province, China. The results show that the Word2Vec-BiGRU-CNN model achieves a high F1-score of 0.3648 for traffic accident duration prediction. This study confirms that the newly established structured features extracted from text data significantly enhance the prediction effects of deep learning algorithms. However, these new features are detrimental to the prediction effects of conventional machine learning algorithms. The results demonstrate that text feature processing and extraction is a complex issue in traffic accident duration prediction. The proposed method uses a BiGRU-CNN architecture to process multi-modal data, including structured data and text data. The model is trained using six defined data modes, which are derived from the dataset. The results show that the BiGRU-CNN model provides the best prediction performance with the highest mean-F1 value when working with text data. The study concludes that the integration of multi-modal data and deep learning models can significantly improve the accuracy of traffic accident duration predictions. The results also highlight the importance of text vectorization in enhancing the performance of deep learning algorithms. The study suggests that future research should explore more advanced data fusion methods to further improve prediction accuracy.