The rapid assessment of the number of casualties after an earthquake requires to consider not only the characteristics of the earthquake fault but also the population distribution in the disaster area and their living environment. Therefore,the assessment of earthquake casualties is a typical complex prediction system. In this article,we constructed an earthquake casualty assessment model for mainland China,based on the Deep Learning Neural Network method with fatalities of 78 earthquake events during 1976-2020. We applied eight critical factors in the model,i.e.,the date,the time,and the season of occurrence,earthquake-affected population and area,the epicenter,and the focal mechanism of the earthquakes. To test the effectiveness of the model,we used nine events,including the 2008 Wenchuan M_S8.0 earthquake and the 2010 Yushu M_S7.1 earthquake,to compare the estimated numbers of casualties with the true values from the investigation. Our results show that the estimation of fatalities for the seven intermediate earthquakes is good enough,with the error of estimated and investigated numbers in the same order. But for the 2010 Yushu earthquake and the 2008 Wenchuan earthquake,the estimated numbers are significantly smaller than the real ones. The seismogenic fault of the 2010 Yushu earthquake,located directly below the Jiegu town,the capital of the Yushu autonomous prefecture,caused more casualties because of high population density. Moreover,it is also because that the fatalities in the 2008 Wenchuan earthquake was caused not only by the earthquake,but also the secondary disasters.