The on-board emergency call system with accurate occupant injury prediction can help rescuers deliver more targeted traffic accident rescue and save more lives. We use machine learning methods to establish, train, and validate a number of classification models that can predict occupant injuries (by determining whether the MAIS (Maximum Abbreviated Injury Scale) level is greater than 2) based on crash data, and ranked the correlation of some factors affecting vehicle occupant injury levels in accidents. The optimal model was selected by the model prediction accuracy, and the Grid Search method was used to optimize the hyper-parameters for the model. The model is based on 2799 two-vehicle collision accident data from NHTSA CISS (The Crash Investigation Sampling System of NHTSA) traffic accident database.The results show that the model achieves high-precision prediction of occupant injury MAIS level (recall rate 0.8718, AUC(Area under Curve) 0.8579) without excluding vehicle model, and the top 8 relevant features given by the model are: lateral speed change, occupant age, longitudinal speed change, seat belt usage, occupant gender, lateral speed change direction, airbag trigger, and longitudinal speed change direction. We believe that this method can be used to complete organ-level post-crash injury prediction after adding more features, which has great potential to improve the efficiency of traffic accident rescue and reduce the casualty rate.