Prediction of insulin resistance using multiple adaptive regression spline in Chinese women.

Abstract

Insulin resistance (IR) is the core for type 2 diabetes and metabolic syndrome. The homeostasis assessment model is a straightforward and practical tool for quantifying insulin resistance (HOMA-IR). Multiple adaptive regression spline (MARS) is a machine learning method used in many research fields but has yet to be applied to estimating HOMA-IR. This study uses MARS to build an equation to estimate HOMA-IR in pre-menopausal Chinese women based on a sample of 4,071 healthy women aged 20-50 with no major diseases and no medication use for blood pressure, blood glucose or blood lipids. Thirty variables were applied to build the HOMA-IR model, including demographic, laboratory, and lifestyle factors. MARS results in smaller prediction errors than traditional multiple linear regression (MLR) methods, and is thus more accurate. The model was established based on key impact factors including waist-hip ratio (WHR), C reactive protein (CRP), uric acid (UA), total bilirubin (TBIL), leukocyte (WBC), serum glutamic oxaloacetic transaminase (GOT), high-density lipoprotein cholesterol (HDL-C), systolic blood pressure (SBP), serum glutamic pyruvic transaminase (GPT), and triglycerides (TG). The equation is as following:HOMA-IR = 6.634 - 1.448MAX(0, 0.833 - WHR) + 10.152MAX(0, WHR - 0.833) - 1.351MAX(0, 0.7 - CRP) - 0.449MAX(0, CRP - 0.7) + 1.062MAX(0, UA - 8.5) + +1.047(MAX(0, 0.83 - TBIL) + 0.681MAX(0, WBC - 11.53) - 0.071MAX(0, 11.53 - WBC) + 0.043MAX(0, 24 - GOT) - 0.017MAX(0, GOT - 24) + 0.021MAX(0, 59 - HDL) - 0.005MAX(0, HDL - 59) - 0.013MAX(0, 141 - SBP) - 0.033MAX(0, 100 - GPT) + 0.013MAX(0, GPT - 100) - 0.004MAX(303 - TG)Results indicate that MARS is a more precise tool than fasting plasma insulin (FPI) levels, and could be used in the daily practice, and further longitudinal studies are warranted.