Unsupervised Machine Learning approach on Metabolic Syndrome patient identification

Abstract

Metabolic syndrome (MetS) represents a complex constellation of metabolic abnormalities, including abdominal obesity, dyslipidemia, hypertension, and impaired glucose metabolism, that significantly elevate the risk of cardiovascular diseases and type 2 diabetes mellitus. Despite its public health relevance, the clinical definition of MetS remains ambiguous due to variations in diagnostic criteria used by a plethora of organizations such as the World Health Organization definition (WHO), the European Group for the Study of Insulin Resistance definition (EGIR), the National Cholesterol Education Program (NCEP) and the International Diabetes Federation (IDF). These vague definitions create discrepancies that complicate patient clustering and segmentation, making it challenging to correctly identify them and develop consistent therapeutic approaches. Additionally, real-world datasets frequently exhibit missing values, further complicating the accurate identification of patient subgroups and the quantification of their associated risks. This thesis addresses and is driven by two major issues in the study of metabolic syndrome: (1) the variability in diagnostic definitions and its impact on patient clustering, and (2) the challenges posed by missing data in real-world clinical datasets. Initially, the four most popular MetS definitions are translated into lines of code to identify and label patients with real data. Afterwards, clustering algorithms and combination of methods are utilized to group patients based on their clinical profiles. To address the issue of missing data, cutoffs are set and data interpolation is applied. Using a dataset of 850 patients, this analysis reveals that differing definitions of MetS result in significant variation in MetS patient identification and that they have to be combined in order to get a robust view, with some individuals probably being still misclassified or excluded entirely depending on the criteria used. Furthermore, the presence of missing values, particularly in variables crucial to the MetS definitions like glucose and waist circumference, disrupts clustering algorithms, leading to biased assessment and non-valid clustering separation. To mitigate these challenges, hybrid methodologies are tested combining unsupervised machine learning techniques, such as K-means and Spectral clustering with principal component analysis (PCA) to handle the dimensionality, with robust missing data handling and exclusion methods. The results demonstrate that the application of different clustering techniques can improve the clustering outcomes and patient classification, providing a clearer understanding of MetS risk stratification. On the other hand, while the data remain unreliable and the MetS definitions vary, no robust classification results are enabled to occur. This work underscores the need for standardizing MetS definitions and addressing missing data to improve clinical outcomes and the precision of metabolic risk management strategies.