Analysis of Interacting Factors Contributing to the Onset of Diabetes Type 2 in the Female Pima Population: A Step towards Understanding the Links

摘要

One of the biggest challenges in building predictive models relates to proper feature selection. As it is well known, good feature selection will result in accurate prediction independently of the classifier. We propose a novel processing pipeline consisting of unsupervised clustering to balance the class prevalence in the training set, which enables the development of stable models with less overfitting. In this paper, we use the proposed pipeline to predict the onset of Type 2 Diabetes (T2D) in the female Pima American Indian population. The Pima population, living near Phoenix, Arizona in the United States, has been known to have the highest T2D prevalence. While it generally remains unclear how various physiological variables and interacting risk factors contribute to the onset of T2D, here for the first time, we quantify the contribution of each of the physiological variables and their interactions to the predicted probability of developing T2D. Our findings in the adult Pima female T2D cohort point to the existence of two distinct population sub-groups, separated primarily by age and the number of pregnancies. For the first time in this dataset, we show several factor interactions: a) subjects with lower BMI (26.0 ± 2.2 vs. 41.5 ± 4.9) experienced a 57-78% reduction in T2D conversion rate at the high plasma glucose (>133.7 mg/dL) and high serum insulin (>168.0 pmol/L) levels respectively, and b) low serum insulin (70.1 ± 16.7 vs. 294.7 ± 136.6 pmol/L) was associated with a 90% lower risk of T2D conversion rate in individuals with high triceps skin thickness (>34 mm), and interestingly c) non-hypotensive (85.6 ± 7.2 vs. 54.3 ± 7.7 mm Hg) women with ≤ 1 pregnancy had 81% lower risk of T2D conversion. With an increasing number of pregnancies, the trend reversed, and non-hypotensive women experienced a 41% higher rate of T2D conversion. Through the evaluation of multiple feature selection methods and classifiers, we show the potential value of the proposed methodology for feature selection and analysis of interacting factors contributing to the onset of T2D in the female Pima Population.