VALIDATION OF CLUSTERING METHODS FOR MEDICAL DATA SETS
Journal Title: Acta HealthMedica - Year 2017, Vol 2, Issue 1
Abstract
Introduction: Data mining techniques have been increasingly applied to medical data in the past decade and are divided into two categories: predictive algorithms such as classification and descriptive algorithms such as clustering and association rule mining (ARM). Clustering means partitioning a data set into a set of clusters in such a way that the samples belonging to the same clusters are similar and those belonging to different clusters are dissimilar. These algorithms are currently used as a preprocessing technique prior to medical data mining and analyzing. However, clustering algorithms have different behaviors depending on the features of the data set. Therefore, in most applications, results of clustering are evaluated in terms of some validity clustering measures. In addition, most of the medical data sets are usually complex, with nonlinear patterns, which are extremely large and difficult to cluster. The aim of this study is to performance analyze several clustering techniques based on the basis of two different classes of clustering quality measures, named internal and stability, over the medical data sets. Methods: In this study, the performance of six common clustering algorithms, including hierarchal clustering, K-means, fuzzy C-means (FCM), self-organizing tree algorithm (SOTA), Diana and partitioning around medoids (PAM), has been examined by using four different medical data sets, which are available on UCI repositories, including the Indian liver patient data set (ILPD), Pima Indians diabetes, breast cancer, and statlog heart disease. The results of clustering methods have been evaluated based on two different classes of clustering quality measures called “internal” and “stability.” Stability measures include average proportion of non-overlap (APN), average distance (AD), average distance between means (ADM), and figure of merit (FOM) and internal measures consist of connectivity, Silhouette and Dunn index. Results: Due to given data sets consisting of two classes of samples, the number of cluster is assumed to two. The evaluation results showed that hierarchical clustering is the best algorithm to cluster all of the current data sets on the basis of internal validity measures. While based on stability measures, a Diana algorithm for ILPD and breast data sets, hierarchical clustering method for heart data set and SOTA for Pima data set outperformed other clustering algorithms. Conclusion: In this study, performance of several clustering techniques on the basis of internal and stability clustering measures over the medical data sets were investigated. The hierarchical clustering and Diana methods revealed better results. The number of clusters is assumed to two, while this parameter can have an effect on performance of different clustering algorithms. Future work will mainly cover the optimal estimation number of clusters and also consider more clustering methods for evaluation
Authors and Affiliations
Azam Orooji, Farzaneh Kermani
Keywords
Related Articles
- EP ID EP350545
- DOI 10.19082/ah116
- Views 108
- Downloads 0
DEVELOPMENT OF A NATIONAL CORE DATA SET FOR THE IRANIAN ICU PATIENT OUTCOME PREDICTION
Introduction: To define a core data set for ICU patients outcome prediction in Iran. This core data set will lead us to design ICU outcome prediction models with the most effective parameters. Methods: A combination of...
GASTRIC, ESOPHAGEAL, AND COLON CANCERS BIOBANK IN NORTHEAST IRAN
Introduction: Gastric, esophageal and colon cancers are prevalent in Iran. Human bio specimens are increasingly utilized as materials in cancer research. Modern, molecular biomedicine is driving a growing demand for exte...
MOLECULAR DOCKING STUDY OF A NUCLEOTIDASE PURIFIED FROM CERASTES CERASTES VENOM: PROSPECT OF USE IN THE TREATMENT OF CD 73 DEFICIENCY
Background: Deficiency in ecto-5′-nucleotidase CD73 alters thromboregulation and affects coronary vascular tone and platelet activation. Previous studies reported that the treatment with soluble 5’-nucleotidase inhibited...
GENETIC EPIDEMIOLOGICAL CHARACTERIZATION OF THE POPULATION OF TLEMCEN (WEST ALGERIA) BY TYPE 1 DIABETES
Background: Type 1 diabetes is a real public health problem, concerning not only its complications and cost, but also its increasing incidence and the fact that it occurs in young people. The objective of this study is t...
SEPHYRES 2: A SYMPTOM CHECKER BASED ON SEMANTIC PSEUDO-FUZZY DIAGNOSTIC MODEL
Introduction: The symptom checkers were designed to help patients and health professionals. In SEPHYRES 1 symptom checker, a new viewpoint of medical ontology and two reasoning strategies were developed based on pain-onl...