Abstract: Based on the five evaluation index data of 70 earthquakes in China from 1990 to 2011 and the number of earthquake casualties, a projection pursuit regression (S-PPR) model based on SMART algorithm was established by using five-fold cross-validation method. The proportion of the sample with the prediction error of less than 0, 1, 2 and 10 people was 41.2%, 61.8%, 75% and 92.6%, respectively. The proportion of the sample with the prediction error of less than 1 person in the four test samples was 75%, which indicated that the S-PPR model had high prediction accuracy. Among the five evaluation indexes, the earthquake magnitude has the most significant impact on the number of earthquake casualties, followed by the population density, by the time of the earthquake, the epicenter intensity, and the ΔL the relatively small impact. For the data of this example, multiple-fold cross-validation method is not adopted, although the BPNN model structure meets the most basic requirements of establishing BPNN model, and the BPNN model with small error can be "pick out" and "high precision" for the training set data as well as the testing set data, the “pick out” BPNN model has no generalization ability and practical value. Moreover, the main problems in the published articles in the process of establishing the BPNN, RFR, SVM, RBFNN and ELM models for earthquake casualty assessment are also analyzed. Compared with MLR and BPNN models established by cross-validation method, S-PPR model has better data fitting ability, generalization ability and robustness, and also expands new methods and techniques for earthquake casualty assessment.