Investigation of a Classification about Time Series Signal Using SOM

Abstract

There are various blood purifications. Plasma Exchange(PE) is one of the blood purifications. A patient's circulating blood volume is changing during the PE. Hence, the medical workers must treat to stabilize the circulating blood volume. They are the CLIT-LINE monitor (CLM) to obtain the hematocrit (Ht) values, because the circulating blood volume is calculated by Ht values. Ht value reflects the count of red cells in the circulating blood volume, and Ht values are changed by the urination. They recognize the patient's condition by the classifying the change of Ht values. However, the classifying the change of Ht values is depended on the experience of medical workers. This approach is very subjective. Hence, our purpose is the investigation of the possibility about the classifying the change of Ht value using the Self Organized feature Map (SOM). We used the Ht value that was obtained in the Tokushima University Hospital. The Ht data was greatly changed by urination. At first, the Ht data is divided into the consecutive grouped patterns. We used recurrent neural network (RNN) to treat the change of Ht value in each group. RNN is effective to treat the feature of time series signals, and it was known that the weight values of RNN have some information about the feature of signals. The seconds, the weight values obtained in each group are used as input data to SOM. As a result of SOM, Ht data were divided into 4 groups. One of them was a group into which Ht values were greatly changed by urination. It is thought that other groups show the state that has changed little by little during the PE. As a conclusion, it was supposed that the possibility about the classifying the change of Ht value using SOM.