Electronic Health Record Driven Prediction for Gestational Diabetes Mellitus in Early Pregnancy

Abstract

Gestational diabetes mellitus (GDM) is conventionally confirmed with oral glucose tolerance test (OGTT) in 24 to 28 weeks of gestation, but it is still uncertain whether it can be predicted with secondary use of electronic health records (EHRs) in early pregnancy. To this purpose, the cost-sensitive hybrid model (CSHM) and five conventional machine learning methods are used to construct the predictive models, capturing the future risks of GDM in the temporally aggregated EHRs. The experimental data sources from a nested case-control study cohort, containing 33,935 gestational women in West China Second Hospital. After data cleaning, 4,378 cases and 50 attributes are stored and collected for the data set. Through selecting the most feasible method, the cost parameter of CSHM is adapted to deal with imbalance of the dataset. In the experiment, 3940 samples are used for training and the rest 438 samples for testing. Although the accuracy of positive samples is barely acceptable (62.16%), the results suggest that the vast majority (98.4%) of those predicted positive instances are real positives. To our knowledge, this is the first study to apply machine learning models with EHRs to predict GDM, which will facilitate personalized medicine in maternal health management in the future.

Figure 1

Prediction model and data processing schematic diagram. In EHRs, the feature vectors were extracted from the characteristics of the first trimester and the class labels from the diagnostic international classification of diseases (ICD-10) codes of OGTT in 24–28 weeks’ gestation. After EHR preprocessing, the experimental data were divided into two subsets in evaluation design. The training set was then modelled using six machine learning techniques and the variants of cost-sensitive hybrid models (CSHM). Five measure metrics of the performance were collected: accuracy; area under the ROC curve (AUC), true positive rates, false positive rates and confidence reports.

Figure 2

Performance of six techniques with cross validation. Bar graphs in (A), (B), (C) and (D) illustrate accuracy, area under ROC curve (AUC), true positive rate (TPR) and false positive rate (FPR) of those six techniques, respectively. Curves in (E) and (F) demonstrate receiver operating characteristic (ROC) for training and testing. LR: logistic regression; NB: naive Bayes; NN: neural network; SVM: support vector machine; CHAID: Chi-square automatic interaction detection Tree; CSHM (1): cost-sensitive hybrid model with cost parameter λ₁=1 (symmetrical costs of misclassification). TPR and FPR are obtained from their confusion matrix.

Figure 3

Performance of CSHM in five cost sensitive contexts with cross validation. Bar graphs in (A), (B), (C) and (D) illustrate accuracy, area under ROC curve (AUC), true positive rate (TPR) and false positive rate (FPR) of CSHM in five cost sensitive contexts, respectively. Curves in (E) and (F) demonstrate receiver operating characteristic (ROC) for training and testing. CSHM (1.5): cost-sensitive hybrid model with cost parameter λ₁ = 1.5 (asymmetrical costs of misclassification). TPR and FPR are obtained from their confusion matrix.

Figure 4

Significance of CSHM comparing with other methods. (A) Significance of CSHM to the algorithms of SVM, LR and NN; (B) significance of CSHM(100) to the other four cost sensitive contexts. (C) Comparison of the results with CSHM and SVM on the experimental data set. T(1): CSHM(1), CSHM model takes the cost parameter λ₁=1. T(1)-LR (or NN, SVM): the true positive rates of CSHM(1) minus those of LR (or NN, SVM). T(100)-T(1)(or T(5), T(10), T(1000)): the true positive rates of CSHM(100) minus those of CSHM(1) (or T(5), T(10), T(1000)). p-value < 0.001 illustrates the significance of those two methods with a two-sided test for difference in AUC.

Figure 5

Confidence reports of six techniques and CSHM in five cost sensitive contexts with cross validation. Bar graphs in (A) and (B) illustrate mean correct and bar graphs in (C) and (D) illustrate mean incorrect of those six techniques and CSHM in five cost sensitive contexts, respectively. Boxplots in (E) and (F) illustrate confidence distributions for training and those in (G) and (H) illustrate confidence distributions for testing of those six techniques and CSHM in five cost sensitive contexts, respectively. Mean correct: mean confidence of correct predictions; mean incorrect: mean confidence of incorrect predictions.