Purpose In Korea, social care services are becoming important according to changes in the demographic structure due to aging. Case management using big data is used to effectively respond to this. The use of big data provides new insights into social service planning, goal setting, and delivery (Gillingham et al. 2017), and serves as an alternative methodology when qualitative methods for accessing dangerous or vulnerable classes are not sufficient or reasonable (Zetino et al. 2019). Recently, machine learning-based prediction models were employed to understand the complex needs in case management and to accurately support the resources that matched them. The purpose of our research is also to develop a machine learning-based prediction model for three big data surveyed by the Social Security Information Service and local governments. Method From the three big data collected from 2019-2020, a categorical variable consisting of four needs (maintaining daily life, economy, health, and living environment) and a categorical variable consisting of five resources (physical health, health management, mental health, psychological sentiment, and housing) were defined as two outcome variables. For independent variables, depending on how we combine the big data, three datasets including multiple variables related to physical and mental health and house structure related variables were constructed. After preprocessing by combining the three sets of independent variables with the two outcome variables, six datasets were created with different sample size(n) and number of variables(p): (1) the three datasets related to the needs - dataset 1(n=15,434, p=176), dataset 2(n=4,075, p=334) and dataset 3(n=600, p=412), (2) the three datasets related to the resources - dataset 4(n=15,434, p=208), dataset 5(n=4,075, p=365) and dataset 6(n=600, p=419). In building the prediction model for the resources as outcome variable, needs-related variables were added to the independent variables. We considered two methods for multi-label classification (Zhang et al. 2013). One method called "Binary relevance" was considered to build machine learning models for each binary category from the four needs and the five resources. The other method called "Chain ensemble classifier" was considered to build machine learning models using the four needs and the five resources as outcome variables. The prediction models included logistic regression, lasso, svm, randomforest, xgboost (Hastie et al. 2009), and they were tuned by 10-folds cross-validation in the train set, and best prediction models were evaluated using F1-score in the test set. The ratio of train set and test set was divided into 3:1. Results and Discussion First, we identified the predictive performance results of each best model for the four needs: Daily life (F1=0.866), economy (F1=0.720), health (F1=0.870), and living environment (F1=0.768). For the resources, the results were physical health (F1=0.893), health management (F1=0.949), mental health (F1=0.553), psychological sentiment (F1=0.772), and housing (F1=0.916), respectively. The high predictive performance of the models was confirmed, and in particular, noticeably high in two types, health management and housing(>0.9). This result shows an opportunity to confirm the possibility of machine learning-based models, which is still rarely applied to national policies and practices, as a decision support tool for case managers. [ABSTRACT FROM AUTHOR]