Combinatorial K-Means Clustering as a Machine Learning Tool Applied to Diabetes Mellitus Type 2

• Published in: International journal of environmental research and public health Vol. 18; no. 4; p. 1919
• Main Authors: Nedyalkova, Miroslava, Madurga, Sergio, Simeonov, Vasil
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI 17.02.2021
• Subjects: Algorithms; Cluster Analysis; Diabetes Mellitus, Type 2 - epidemiology; Humans; Machine Learning; k-means; descriptors; underlying diseases; classification; diabetes mellitus type 2; multivariate statistics
• ISSN: 1660-4601; 1661-7827; 1660-4601
• DOI: 10.3390/ijerph18041919

A new original procedure based on k-means clustering is designed to find the most appropriate clinical variables able to efficiently separate into groups similar patients diagnosed with diabetes mellitus type 2 (DMT2) and underlying diseases (arterial hypertonia (AH), ischemic heart disease (CHD), diabetic polyneuropathy (DPNP), and diabetic microangiopathy (DMA)). Clustering is a machine learning tool for discovering structures in datasets. Clustering has been proven to be efficient for pattern recognition based on clinical records. The considered combinatorial k-means procedure explores all possible k-means clustering with a determined number of descriptors and groups. The predetermined conditions for the partitioning were as follows: every single group of patients included patients with DMT2 and one of the underlying diseases; each subgroup formed in such a way was subject to partitioning into three patterns (good health status, medium health status, and degenerated health status); optimal descriptors for each disease and groups. The selection of the best clustering is obtained through the parameter called global variance, defined as the sum of all variance values of all clinical variables of all the clusters. The best clinical parameters are found by minimizing this global variance. This methodology has to identify a set of variables that are assumed to separate each underlying disease efficiently in three different subgroups of patients. The hierarchical clustering obtained for these four underlying diseases could be used to build groups of patients with correlated clinical data. The proposed methodology gives surmised results from complex data based on a relationship with the health status of the group and draws a picture of the prediction rate of the ongoing health status.