Clustering Strategy Validity Analysis On Ambulatory Electrocardiogram

Clustering is an unsupervised learning algorithm. It divided the data set into different classes by certain criteria. The goal is to make the objects within the class as similar as possible, and make the differences among classes as far as possible. In cluster analysis, is the clustering algorithm suitable for a given data set? Can the clustering results reflect the inherent structure? These questions introduced the evaluation of clustering results.Validity methods of traditional clustering are mostly used in low-dimensional data clustering results, and good results were achieved. The study selected AECG (Ambulatory Electrocardiogram) as example. Since AECG waveforms are high-dimension data, traditional validity methods show limited ability. Based on the analysis of biological background of AECG and clustering validity methods, the thesis studied internal criteria and relative criteria on validity of AECG cluster results.This thesis proposes improved FOM (Figure of Merit) method though analysis the characteristics of AECG waveform. FOM method is a classic method of internal validity, but the method reflected the differences by Euclidean distance within the class, is not suitable for evaluated the clustering results of AECG waveform. Hausdorff distance is a minimax distance, it not need to create the relationship between everyone points, just calculated the degree of similarity of two sets. So, an improved validation method is presented in the thesis based on the FOM method, which computing Hausdorff distance through the weighted sum of ECG segments. The experiments using MIT-BIH standard database of AECG data, compared with traditional method FOM, the proposed one has more validity.Determine the optimal number of clusters is one of the difficulties in clustering research. This thesis proposed a relative method based on gravity concept. Create the validity function from the compactness within the class and the separation between classes, in order to determine the optimal cluster number of categories. The experiments using MIT-BIH standard database of ECG data, compared with traditional methods SD method and DB method, the proposed one has clear advantage to determine the validity clustering.Finally, for the proposed methods, use the standard MTI-BIH database of AECG waveform design experiments. The experiments results can guide users to select clustering algorithm which appropriate data set, and can the best clustering number of classes.