A secondary analysis of a large community data set from the 2008 national study of Korean elderly was conducted to provide accurate information regarding the characteristics of the elderly at risk of malnutrition in Korea. The Institutional Review Board gave prior approval to the study. The sample consisted of 15,146 adults over 60 years old and community dwelling. The ten items of the 'Determine Your Nutritional Health Checklist' were used to access dietary intake and meal patterns. The tool was developed by the Nutritional Screening Initiative to screen nutritional risk in the elderly population. A summed score of 0-2 indicates good nutritional status, 3-5 moderate risks, and 6 or more high risk [15]. In this study, elderly people with summed scores of 3 or more were classified as being at risk of malnutrition as they pose problems for malnutrition.

Figure 1 presents the data mining process that started with the selection of variables. The target variable was malnutrition, and the 52 input variables covered all aspects of elderly malnutrition, including general characteristics, family & social relationship, economic status, health status, health behavior, functional status, leisure & social activity, quality of life, and living environment (Figure 2). The ratio between the training data and test data was 7:3 by random division of the full data set to which four tree-structured models, including C5.0, CHAID, QUEST, and classification and regression (C&R) Tree, were applied. In an assessment of model performance, analysis node was used to calculate the percentage of correct and wrong classification of each model, and then the correct classification rates of the models were compared to select the best model. In this study, the C&R Tree showed the best predictive performance, and it was selected to draw the final decision support tree [16]. A decision tree is a visual representation of obtained knowledge using a tree form in which each branch represents an output of the test. This decision tree enables easy understanding and interpretation of data as the nodes and branches are organized hierarchically. It generates reliable knowledge, promoting accuracy in clinical decision-making [17]. The C&R Tree node is a tree-based classification and prediction method. The C&R Tree approach provides a very simple representation that displays accumulated knowledge well. This method allows the tree to grow large before pruning, which results in smaller trees, and this ensures better cross-validation properties of decision tree modeling [18]. The recursive partitioning process splits the records into segments with similar output field values, and two subgroups are defined by the first split. The split ends when one of stopping criteria is triggered [19].

The default experimental parameters of Clementine ver. 12.0 (SPSS Inc., Chicago, IL, USA) were utilized to construct the decision tree model with the usage of the C&R Tree component in this study. SPSS ver. 12.0 for Windows (SPSS Inc.) was used to perform the statistical analysis in which important variables were chosen using the feature selection method. Among Pearson, Likelihood ratio, Cramer's V, and Lambda-a, Pearson correlation analysis was selected to prioritize the meaningful variables [20].

A total of 15,146 elderly subjects were included in this study, and the sample consisted of 6,185 males (40.8%) and 8,961 females (59.2%). Of the total participants, 7,769 (51.3%) had normal nutritional status while 7,377 (48.7%) had poor nutritional status. Compared with elderly subjects showing good nutritional status, the mean age of elderly subjects at risk of malnutrition was higher (Table 1). About three-fourths of older people living alone were categorized as having poor nutritional status, and there was a greater proportion of elderly subjects with good nutritional status in the higher monthly income group. Apart from 54 participants (0.4%), the majority of participants had been married at least once. While malnourished elderly subjects comprised about one third of the married group, the proportion of malnourished elderly subjects was much greater in the separated or living alone groups, accounting for 73.5% and 65.2%, respectively.

As a result of the feature selection process, 48 important attributes were identified to be important variables of a total of 52 input variables shown in Figure 2. Seven variables including chewing ability, level of life satisfaction, depression status, health status, number of diagnosed diseases, difficulties in daily activities caused by pain, subjective economic state, and monthly income were identified to be most significant variables.

The highest percentage of correct prediction in training data was reported in C5.0; however, the C&R Tree model showed relatively higher percentages of correct prediction in both training and test data with 78.10% and 80.95%, respectively (Table 2).

Figure 2 presents seven finalized input variables that were selected during feature selection. The final decision support model was designed using the C&R Tree algorithm and is shown in Figure 3.

In this study, a total of six rules were identified to be associated with the risk of malnutrition in the elderly. The six decision rules were the following: 1) good and very good chewing ability, depressed, level of life satisfaction less than 2.89; 2) good and very good chewing ability, depressed, level of life satisfaction above 2.89, severe difficulty in daily activity caused by pain; 3) low chewing ability, living with spouse or children; low chewing ability, living with spouse or children, depressed; 5) low chewing ability, living with spouse or children, not depressed, number of disease above 2.5; 6) low chewing ability, living with spouse or children, not depressed, number of disease less than 2.5, subjective income status is poor.

The traditional statistical method and chi-square test were used to find out whether there are significant differences between the elderly at risk of malnutrition and normal subjects in the seven final variables identified by the C&R Tree algorithm. All seven variables showed statistical significance with a p-value of <0.05 (Table 3). Chewing ability, in particular, showed the biggest differences between well-nourished participants and malnourished participants. The majority of participants with low to moderate chewing ability had poor nutrition status (84.1% and 70.2%, respectively) while eight out of ten participants with either good or very good chewing ability were well-nourished. In addition higher proportions of individuals with depression, low life satisfaction, many difficulties in daily activities and poorly perceived economic status were found to be at risk of malnutrition.