Identification of insulin resistance in Asian Indian adolescents: classification and regression tree (CART) and logistic regression based classification rules

• Published in: Clinical endocrinology (Oxford) Vol. 70; no. 5; pp. 717 - 724
• Main Authors: Goel, Ruchika, Misra, Anoop, Kondal, Dimple, Pandey, Ravindra M., Vikram, Naval K., Wasir, Jasjeet S., Dhingra, Vibha, Luthra, Kalpana
• Format: Journal Article
• Language: English
• Published: Oxford, UK Blackwell Publishing Ltd 01.05.2009; Blackwell
• Subjects: Adiposity; Adolescent; Adult; Asian Continental Ancestry Group; Biological and medical sciences; Classification - methods; Cross-Sectional Studies; Decision Trees; Diabetes Mellitus, Type 2 - etiology; Endocrinopathies; Female; Fundamental and applied biological sciences. Psychology; Humans; India; Insulin Resistance - ethnology; Logistic Models; Male; Medical sciences; Regression Analysis; Risk Factors; ROC Curve; Skinfold Thickness; Vertebrates: endocrinology; Young Adult; Asia; India; Endocrinopathy; Human; Pancreatic hormone; Asiatic; Metabolic diseases; Identification; Insulin; Logistic regression; Target tissue resistance; Adolescent; Classification; Tree; Indian; Insulin resistance; Endocrinology
• DOI: 10.1111/j.1365-2265.2008.03409.x
• ISSN: 0300-0664; 1365-2265

Summary Objective  Biochemical measures for assessment of insulin resistance are not cost‐effective in resource‐constrained developing countries. Using classification and regression tree (CART) and multivariate logistic regression, we aimed to develop simple predictive decision models based on routine clinical and biochemical parameters to predict insulin resistance in apparently healthy Asian Indian adolescents. Design  Community based cross‐sectional study. Subjects and patients  Data of apparently healthy 793 adolescents (aged 14–19 years) were used for analysis. WHO's multistage cluster sampling design was used for data collection. Methods and measurements  Homeostasis Model of Assessment value > 75th centile was used as cut‐off for defining the main outcome variable insulin resistance. CART was used to develop the decision tree models and multivariate logistic regression used to develop the clinical prediction score. Results  Three classification trees and an equation for prediction score were developed and internally validated. The three decision trees were termed as CART I, CART II and CART III, respectively. CART I based on anthropometric parameters alone has sensitivity 88·2%, specificity 50·1% and area under receiver operating characteristic curve (aROC) 77·8%. CART II based on anthropometric and routine biochemical parameters has sensitivity 94·5%, specificity 38·3% and aROC 73·6%. CART III based on all anthropometric, biochemical and clinical parameters together has sensitivity 70·7%, specificity 79·2% and aROC 77·4%. Prediction score for insulin resistance = 1 × (waist circumference) + 1·1 × (percentage body fat) + 1·6 × (triceps skin‐fold thickness) – 1·9 × (gender). A score cut‐off of > 0 (using values marked for each) was a marker of insulin resistance in the study population (sensitivity 82·4%, specificity 56·7%, and aROC 73·4%). Conclusion  These simple and cost‐effective classification rules may be used to predict insulin resistance and implement population based preventive interventions in Asian Indian adolescents.