【Objective】To improve the accuracy of predicting the severity of traffic accidents and clarify the key in fluencing factors of traffic accident severity.【Method】Fourteen factors related to people, vehicles, roads, and envi ronment that affect traffic accident severity were selected from 885 road traffic accident case data, and a traffic acci dent classification prediction model was constructed using a convolutional neural network with channel attention (CA-CNN). On this basis, the significant influencing factors of traffic accident severity were analyzed using the method of marginal utility.【Result】The results show that compared with the 6 prediction models such as convolu tional neural network (CNN), random forest (RF), NaiveBayes, logistics regression (Logistics), decision table (De cision_table), and bagging algorithm (Bagging), the convolutional neural network model with channel attention fu sion has better overall prediction performance in terms of accuracy, precision, and recall, etc. Among the influenc ing factors of traffic accident severity, the season, whether it is a workday, road type, accident type, illegal lane change, failure to yield, and improper braking have significant effects (p≤0.05).【Conclusion】CA-CNN is an effec tive traffic accident severity prediction model, and the analysis results are helpful in reducing the incidence and se verity of traffic accidents.