A combined strategy of feature selection and machine learning to identify predictors of prediabetes

Abstract

Objective: To identify predictors of prediabetes using feature selection and machine learning on a nationally representative sample of the US population.

Materials and methods: We analyzed n = 6346 men and women enrolled in the National Health and Nutrition Examination Survey 2013-2014. Prediabetes was defined using American Diabetes Association guidelines. The sample was randomly partitioned to training (n = 3174) and internal validation (n = 3172) sets. Feature selection algorithms were run on training data containing 156 preselected exposure variables. Four machine learning algorithms were applied on 46 exposure variables in original and resampled training datasets built using 4 resampling methods. Predictive models were tested on internal validation data (n = 3172) and external validation data (n = 3000) prepared from National Health and Nutrition Examination Survey 2011-2012. Model performance was evaluated using area under the receiver operating characteristic curve (AUROC). Predictors were assessed by odds ratios in logistic models and variable importance in others. The Centers for Disease Control (CDC) prediabetes screening tool was the benchmark to compare model performance.

Results: Prediabetes prevalence was 23.43%. The CDC prediabetes screening tool produced 64.40% AUROC. Seven optimal (≥ 70% AUROC) models identified 25 predictors including 4 potentially novel associations; 20 by both logistic and other nonlinear/ensemble models and 5 solely by the latter. All optimal models outperformed the CDC prediabetes screening tool (P < 0.05).

Discussion: Combined use of feature selection and machine learning increased predictive performance outperforming the recommended screening tool. A range of predictors of prediabetes was identified.

Conclusion: This work demonstrated the value of combining feature selection with machine learning to identify a wide range of predictors that could enhance prediabetes prediction and clinical decision-making.

Keywords: NHANES; feature selection; machine learning; prediabetes; predictors.

Figure 1.

Feature selection using Boruta algorithm: Variable importance plot. Default functions of the “Boruta” R package were used; feature importance measure = mean decrease accuracy, maximal number of random forest runs = 100. Red, yellow, green, and blue boxplots represent Z scores of rejected, tentative, confirmed and shadow attributes respectively. Shadow (minimum, mean, and maximum) features are reference points for deciding which attributes are truly important and these values are generated by the algorithm via shuffling values of the original attributes. Variables extracted from the 20 confirmed and the 10 tentative features selected by the “Boruta” algorithm are given in Table 1. (shadowMin=Minimum shadow score, cpk=creatine phosphokinase, psoriasis=diagnosed psoriasis, milk=milk consumption, mi=diagnosed heart attack, hepc=hepatitis C, basop=basophil count, copd=diagnosed chronic obstructive pulmonary disease, ocp=oral contraceptive use, ldh=lactate dehydrogenase, healthdev=self-rated health trend, wbc=white cell count, citizen=citizenship status, gdm=gestational diabetes, cuttingsalt=reducing salt intake, edu=education, rbc=red cell count, armc=arm circumference)

Figure 2.

Feature selection using recursive feature elimination. A random forest classifier with two-fold cross-validation was specified with other default functions of the “caret” package in R to extract features via recursive feature elimination. Variables extracted from the 30 most important features selected by the recursive feature elimination algorithm are given in Table 1.