Characterisation of distinct polycystic ovary syndrome subtypes by cluster and principal component analyses

Kharis A Burns^1,2,3*Alexander W Stuckey^4,5Scott G Wilson^6,7,8Gerald F Watts^1,9Bronwyn G A Stuckey^1,10,7

• ¹Medical School, University of Western Australia, Perth, Australia
• ²Department of Endocrinology and Diabetes, Royal Perth Hospital, Perth, Western Australia, Australia
• ³SIr Charles Gairdner Osborne Park Health Care Group, Perth, Australia
• ⁴School of Agriculture and Environment, University of Western Australia, Perth, Western Australia, Australia
• ⁵Pink Lake Analytics, Nedlands, Western Australia, Australia
• ⁶School of Biomedical Sciences, University of Western Australia, Perth, Western Australia, Australia
• ⁷Department of Endocrinology and Diabetes, Sir Charles Gairdner Hospital, Perth, Western Australia, Australia
• ⁸Department of Twin Research and Genetic Epidemiology, King's College London, London, England, United Kingdom
• ⁹Lipid Disorders Clinic, Department of Cardiology, Royal Perth Hospital, Perth, Western Australia, Australia
• ¹⁰Keogh Institute for Medical Research (KIMR), Nedlands, Western Australia, Australia

Polycystic ovary syndrome (PCOS) is a common though clinically heterogeneous condition. This study explores PCOS subtypes using two orthogonal statistical analyses of biochemical and anthropometric data. Unsupervised hierarchical cluster analysis and principal component analysis (PCA) of hormonal and metabolic parameters were performed in a cohort of PCOS-affected women, diagnosed based on NIH criteria. Data collected included body mass index (BMI), blood pressure (BP), fasting insulin and glucose (HOMA-IR), gonadotropins, androgens and lipids. Subtypes were explored using unsupervised hierarchical cluster analysis grouping both phenotypic variables and patients into clusters. PCA resolved correlated variables (excluding BMI) into independent factors, with the influence of BMI on the components then explored. 1035 women with PCOS were included in the study, with 975 assessed using cluster analysis and PCA. Two main clusters of variables were evident, one characterized by BP, BMI, HOMA-IR and lipids (triglycerides/cholesterol/LDL), and a second by LH:FSH, androgens, SHBG and HDL). Three separate patient clusters emerged: Cluster A (29.6%% of women) showed higher BP, BMI, HOMA-IR and lipids (triglycerides/cholesterol/LDL), and lower LH:FSH, SHBG and HDL. Cluster C 43.3%) showed lower BP, BMI, HOMA-IR, triglycerides, testosterone, FAI, and higher LH:FSH, DHEAS, androstenedione, 17-hydroxyprogesterone, SHBG and HDL . Cluster B (30.0%) was intermediate. Two components aligned with the cluster analysis: Principal component (PC)1, including HOMA-IR, systolic and diastolic BP, triglycerides, LDL, FAI, and SHBG, was positively correlated with BMI (R2=0.32, p-value<0.0001) and aligned with cluster A. PC2, influenced by testosterone, LH:FSH, FAI, DHEAS, androstenedione and 17-hydroxyprogesterone, with loadings in the opposite direction from LDL and cholesterol, aligned with cluster C, with little relationship with BMI (R2=0.0067, p-value=0.0107). Different metabolic and reproductive PCOS subtypes are evident. Androstenedione and 17-hydroxyprogesterone are important in the reproductive phenotype, highlighting the importance of these hormones in diagnosis and subtype identification and emphasising their significance in understanding PCOS biology as a predominantly hyperandrogenic disorder. BMI influences and exacerbates the metabolic subtype; in the reproductive group and in lean/normal BMI patients there is little relationship between weight and other PCOS-related characteristics. Accordingly, traditional treatment paradigms cannot be generalised to all women and these subtypes may ultimately be viewed as separate disorders.