New occupant safety systems, which adapt their behavior to the severity of an accident, may improve vehicle safety. We present an approach to learn a prediction model, which estimates crash severity prior to collision. Based on accident parameters acquired with precrash sensors, the learned model categorizes impending accidents into one of multiple severity classes. Besides describing an automatic labeling system for data preparation, we investigate the performance of different classifiers. Results on simulation data demonstrate a classification performance of 84% correctly classified test cases. We discuss a potential application and finish with ideas for classification improvement.