Application of machine learning algorithm for predicting gestational diabetes mellitus in early pregnancy^†

Pregnant women who attended medical institutions for an antenatal examination from November 2017 to August 2018 were selected for this study. Inclusion criteria were as follows: positivity on the human chorionic gonadotropin (HCG) test, agreeing to participate, and providing written informed consent. The exclusion criterion was the existence of diabetes before pregnancy.

Through a literature review, the following indicators related to gestational diabetes were identified: physical examination results (age, weight, blood pressure, blood type, etc.); medical history (hypertension, diabetes, heart disease, histories of abnormal pregnancy, polycystic ovary syndrome (PCOS), etc.); personal history (birth weight of the pregnant woman at the time of her own birth, her mother's weight at birth, her mother's parity at birth, smoker or nonsmoker, etc.); family history (immediate family suffering from diabetes or not, etc.); specialist examination results (fundal height, abdomen circumference, gravidity, parity, etc.); and laboratory indicators (amniotic fluid index/depth, blood glucose, routine blood tests, routine urine tests, blood biochemical infection indicators, glycated hemoglobin, etc.).

Investigators with identical training conducted a face-to-face questionnaire survey of participants and reviewed the pregnancy-related medical records to obtain the indicators related to GDM.

After data collection, some indicators were merged or deleted according to the suggestions of obstetric experts and a statistician. Pre-pregnancy BMI was calculated as follows:

BMI = weight (kg)/height (m)². Pregnancy weight gain (WG) was also calculated as follows: WG (kg) = birth weight − pre-pregnancy weight.

Some laboratory indicators (e.g., amniotic fluid index/depth, routine blood tests, routine urine tests, blood biochemistry, infection indicators, and glycated hemoglobin) were also deleted due to lower acquisition rates among the subjects.

All categorical variables were treated as dichotomous variables (0/1). The output variable was based on whether or not diabetes was diagnosed during 24–28 gestational weeks (yes = 1, no = 0). A total of 67 variables were included in our study as model input according to the suggestions of obstetric experts and a statistician (Table 1).

Based on Python, each model was established using a random forest regression algorithm. Random forest is a combinatorial classifier that is based on statistical theory, with the characteristics of selecting data randomly and feature ranking, and is considered suitable to the feature ranking and selection of medical data.²⁰

To systematically train the model and evaluate its accuracy, this study used the train_test_split function to randomly divide the data into 2 categories, namely a 70% training data set and a 30% testing data set. The training data were used to train the model, and the test data were used to validate it, the two of which were performed separately.

It was found that the performance of the prediction models trained using the different sets of indicators or data sets with different ratios of negative/positive cases differed. In this study, 3 datasets (F1, F2, and F3) were established and then compared [see Technical Route (Figure 1)].

Figure 1

The study was approved by the medical ethics committee of the study principal investigator's hospital. All patients were informed about the purposes and methods of the study during the recruitment. Participation was voluntary, and patients could refuse to participate or withdraw at any time. Each participant provided written informed consent. The patients’ IDs were used instead of their names throughout the study.

A total of 4806 analyzable data from 1625 pregnant women were collected. Among these, 3265 samples with all 67 indicators were used to establish data set F1, including 326 cases of GDM and 2939 normal cases. Moreover, 4806 samples with 38 identical indicators were used to establish data set F2, including 1867 cases of GDM and 2939 normal cases.

To explore the impact of sacrificial indicators on GDM prediction, the F3 data set was established using 3265 samples (F1) with 38 indicators (F2), including 326 cases of GDM and 2939 normal cases.

A total of 3 prediction models of GDM were established for the different data sets by using a random forest regression algorithm. The results of the prediction models are shown in Table 2. The accuracy curves for the training set and test set of the 3 prediction models are shown in Figures 2–4.

Figure 2

Figure 3

Figure 4

Confusion matrixes from the prediction results of the F1, F2, and F3 models are shown in Tables 3–5, respectively.

There are 67 input variables in data set F1, 58 of which were included in the random forest model. The 9 variables that had zero weight are as follows: thyroid disease, kidney disease, congenital spinal bifida, benign tumor, past surgical history, multiple pregnancy history, HIV, syphilis, and human papillomavirus (HPV). The weights of the other variables are shown in Figure 5.

Figure 5

There are 38 input variables in data set F2, 37 of which were enrolled in the random forest model, while 1 variable had zero weight (congenital spinal bifida). Sacrifice indicators are indicators of dietary habits and exercise habits. The weights of the top 20 variables are shown in Figure 6.

Figure 6

The F3 data set was established using 3265 samples (F1) with 38 input variables (F2), 31 of which were included in the random forest model. A total of 7 variables had zero weight, and they are as follows: kidney disease, congenital spinal bifida, benign tumor, past surgical history, HIV, syphilis, and HPV. The other variables are shown in Figure 7.

Figure 7

For more information and to obtain the most complete indicator information, 3265 samples with 67 indicators were enrolled in data set F1. The overall prediction accuracy of the F1 model training with the random forest was 93.10% and area under the receiver operating characteristic curve (AUC) was 0.66 (Table 2), indicating the model's good performance. However, the confusion matrix (Table 3) showed that the high accuracy was derived from 735 negative cases for which the prediction was correct in all cases; in contrast, the prediction was correct in only 26 out of 82 positive cases, with too many false-negative results. The low accuracy of 93.10% and the low recognition rate of the model for positive cases may have been related to the large disparity in the numbers of GDM-positive/negative cases (3269/293) in the F1 model.

To utilize all of the available data for the analysis, 4806 samples with 38 indicators were enrolled in data set F2. The numbers of positive/negative cases of GDM increased (1867/2939), but some indicators were excluded, mainly those related to dietary and exercise habits (Table 1). The prediction accuracy of the F2 model was 88.70% and AUC was 0.87, indicating stable and reliable performance. The confusion matrix (Table 4) showed that the prediction was correct in 695 of 735 negative cases, while it was correct in 371 of 467 positive cases. The overall accuracy was thus 88.69%, indicating that the F2 model was more stable and reliable than the F1 model.

To explore the impact of sacrificial indicators on GDM prediction, the F3 data set was established using 3265 samples (F1) with 38 indicators (F2). The overall prediction accuracy of the F3 model was 91.60% and AUC was 0.58. The confusion matrix (Table 5) showed that the prediction was accurate in all 735 negative cases, but in only 13 out of 82 positive cases. The prediction accuracy of positive cases was only 15.85%, which was worse than those of both the F1 and F2 models. This indicated that the sacrifice indicators related to dietary and exercise habits would reduce the predictive performance of the model.

A total of 58 variables obtained from the F1 model were correlated with the incidence of GDM, the top 20 of which were as follows in this model: BMI, coarse grain intake, bean product intake, porridge intake, mother's weight, filling food intake, fruit intake, sweet food intake, weight gain, age, light vegetable intake, walking time (min/week), mushroom intake, dark vegetable intake, sedentary time (min/day), flour food intake, gravidity history, light physical activity, body weight at birth, and milk intake (Figure 5).

Moreover, 37 variables obtained from the F2 model were correlated with the incidence of GDM, the top 20 of which were as follows: fasting glucose, colpomycosis, BMI, uterine height, abdominal circumference, mother's weight, weight gain, body weight at birth, age, family history of diabetes, systolic pressure, diastolic pressure, gravidity history, gynecological diseases, PCOS, number of births by mother, regular menstruation, kidney disease, parity history, and negative reproductive history.

A total of 31 variables obtained from the F3 model were correlated with the incidence of GDM, the top 20 of which were as follows: BMI, age, weight gain, mother's weight, abdominal circumference, fasting glucose, body weight at birth, systolic pressure, diastolic pressure, uterine height, gravidity history, parity history, number of births by mother, family history of diabetes, blood type B, blood type A, regular menstruation, negative reproductive history, blood type AB, and hepatitis.

Among these top 20 variables in the 3 models, there were 6 overlapping variables (BMI, mother's weight, weight gain, age and gravidity history, and body weight at birth), which also had similar rankings.