A comprehensive survey of child abuse by the Ministry of Health, Labor and Wel- fare of Japan (2019) identified 420 potential risk assessment items. However, using all of these assessment items at a child guidance center would overwhelm their capacity and be thus unrealistic. This study aims to select essential assessment items so that they are usable in actual practice, while maintaining its predictive validity. Here, we used the Random Forests algorithm and predicted classifications to identify the need for child protection by child guidance centers or referral to them from child welfare facili- ties of each municipality. We selected the top 30 items adopted by feature importance in the algorithm from the items evaluated to be easily acquirable in the initial action (50 points or more). The model maintained a moderate level of accuracy 0.783 and AUC-ROC 0.900.
