ABSTRACT
In this paper, we analyzed the relation between urination time and other care records using care records data in order to develop a model that predicts urination time for incontinence prevention. The purpose of this study is to predict urination using care records without sensors. Since the conventional method for predicting urination using care records has a limited number of features, we aim to increase the number of features for prediction and improve accuracy by conducting routine analysis of urination. In the first analysis, we visualized a heatmap of the correlation coefficients for each time period in order to examine the time interval pattern of urination. By this analysis, the interval of urination at each time was analyzed for each subject. In the second analysis, we compared records for the hour before and after urination to see if there was a urination routine. In the first analysis, we found individual differences among the subjects and routine of urination intervals. In the second analysis, it was confirmed that the three subjects had in common a high rate of going to urinate before change dressing assistance. In addition, the present experiment was conducted to predict whether there would be a record of urination one hour later in each of the three subjects’ care records. With 75% of the data as training data and 25% as test data, training and testing were conducted for each individual, and features based on the analysis were used for prediction with random forests, resulting in improved prediction accuracy compared to using only features in the conventional method.
