Traffic Incidents such as crashes and vehicle break down have significant impacts on urban expressway operation. With a well-developed incident duration prediction model, the roadside service and operational efficiency of urban expressways could be improved. In this study, instead of utilizing frequently adopted decision tree and survival analysis method to establish the incident duration analysis model, random survival forests model is employed. The random survival forests model can not only overcome the disadvantage of over-fitting problems of decision tree algorithm, but also break through the limitation of restrictive assumptions and solve the problem of identifying interaction of the covariates in traditional survival analysis. This study is conducted based on traffic incident data of Shanghai urban expressways. The traffic incident data is combined with the road geometry data, traffic operation data, and weather condition information; where 80% data is used as training dataset and the remaining 20% as testing dataset. The results show that incident type, length of road, location, remained lane number and traffic volume have significant impacts on incident duration; and the prediction results based on testing dataset indicate that the random survival forests modelis more accurate than random forests model.