Construction and validation of a line chart for gestational diabetes mellitus based on clinical indicators

Abstract

Background

Gestational diabetes mellitus (GDM) is a common complication of mid–to-late pregnancy. Here, we constructed a predictive model for GDM based on a combination of clinical characteristics and relevant serum markers.

Methods

Data from full-term singleton vaginal deliveries from January 2022 to January 2023 were retrospectively collected from the obstetrics department. The data collected were segregated and assigned to training, validation, and external test sets. Maternal demographic characteristics, living and working habits, and haematological indicators, such as liver function and lipids were collected using a questionnaire designed for the study. The “rms” package in R was used to explore GDM-associated factors through stepwise regression at P < 0.05. A predictive model was developed based on the results of multifactorial logistic regression analysis. We then evaluated the differentiation of the column-line graphical model and performed internal and external validation. To assess the accuracy of the bar graphical model, we plotted calibration and decision curves.

Results

Data from 265 pregnant women were included in the training and internal validation sets, and data from 113 pregnant women were included in the external validation set. The logistic regression algorithm screened 8 indicators as predictors. A prediction model was constructed with ALT, TBA, TC, and TG levels while considering whether GDM affects appetite, the husband– wife relationship, family history, and parental relationships as predictors. The Hosmer–Lemeshow goodness-of-fit test revealed that the chi-square values for the modelling, internal validation, and external validation groups (χ² = 5.964, 3.249, and 12.182, respectively) were all P > 0.05. The ROC curve AUCs for the three groups were 0.93 (95% CI: 0.89–0.97), 0.72 (95% CI: 0.62–0.81), and 0.68 (95% CI: 0.53–0.83), respectively.

Conclusion

In this study, a GDM prediction model was constructed to achieve high performance in GDM risk prediction based on routine obstetric tests and information.

Background

Gestational diabetes mellitus (GDM) is a transient form of diabetes mellitus characterized by varying degrees of impaired glucose tolerance that occurs or is first detected during pregnancy. GDM is a common complication of pregnancy. The global standardized prevalence rate of GDM is 14.0%, with the highest standardized prevalence rates of more than 20% reported in the Middle East, North Africa, and Southeast Asia, followed by the Western Pacific and Africa. In addition to geographic factors, economic factors influence GDM prevalence, with the standardized prevalence rates of GDM being higher in higher-income countries [1, 2]. GDM is associated with a significantly increased risk of secondary diabetes and cardiovascular disease in mothers and of macrosomia, neonatal respiratory distress syndrome, and neonatal jaundice in infants [3,4,5,6]. Therefore, GDM is a major risk factor that is detrimental to the health of mothers and babies.

However, the primary causes and pathophysiological mechanisms of GDM have not been fully elucidated. Metabolic disorders, hormonal changes, and dietary habits during pregnancy may be key factors underlying GDM development [7]. In a prospective study, women with GDM had significantly higher liver enzyme levels than those without GDM. One study reported that single or total elevations in liver enzymes in early pregnancy can predict GDM development, with gamma-glutamyltransferase (GGT), alkaline phosphatase, and aspartate aminotransferase being the three most crucial independent risk factors for GDM [8]. Another prospective study conducted in pregnant Chinese women revealed that elevated liver enzymes and a hepatic steatosis index (HSI, a reliable biomarker of nonalcoholic fatty liver disease) were associated with a greater risk of GDM in early pregnancy, even within the normal range. Altered lipid metabolism is a primary mediator of the association between HIS and GDM [9]. The triglyceride glucose index (TyG) and triglyceride (TG)/high-density lipoprotein (HDL) ratio in early pregnancy are known to aid in the identification of pregnant women at risk of GDM, which may facilitate the planning and implementation of early and effective interventions for improving prognosis. The TyG index has demonstrated a predictive ability superior to that of TG/HDL [10]. However, predictive models for GDM that are based on a combination of liver function and lipid levels are unavailable.

Therefore, this study investigated a predictive model of GDM based on a combination of liver function and lipid levels to provide a theoretical basis and clinical reference for preventing and treating GDM and its related complications.

Materials and methods

General information

The study was approved by the ethics committee of our hospital [CZSFYLL2022 No. 004.], and the study subjects signed an informed consent form. From January 2022 to January 2023, data from pregnant women with GDM who had full-term singleton normal vaginal deliveries at the Obstetrics Department of Changzhi Maternal and Child Health Hospital were retrospectively collected. The data from the Changzhi Maternal and Child Health Hospital were segregated into training and internal validation sets. Data from pregnant women collected from Changzhi People’s Hospital were used as the external validation set. The study included data from patients (1) who were diagnosed with GDM when their glucose levels reached or exceeded the standard criteria, that is, an oral 75 g glucose tolerance test and 1- and 2-h blood glucose diagnostic cut-off values of 5.1, 10.0, and 8.5 mmol/L after fasting and meal, respectively; (2) who had singleton pregnancies; and (3) whose complete obstetric examination data were available. The data of patients who (1) had cardiac, hepatic, or renal dysfunctions prior to delivery; (2) had combined intrauterine infections; (3) had type 1 or type 2 diabetes mellitus or complicated polycystic ovary syndrome before pregnancy; and (4) had other pregnancy complications (hypertension, cardiovascular disease, thyroid disease, or liver disease) were excluded.

We designed a questionnaire to collect data related to the demographic characteristics and life and work habits of the mothers, including age, family history, height, weight, weight gain, systolic blood pressure, diastolic blood pressure, heart rate, body mass index, per capita income of the family, educational level, husband–wife relationship, and parental relationship. Questions about whether the mothers were primiparous, their eating habits were affected in early pregnancy, they were working during pregnancy, they were exercising in the first 3 months of pregnancy, they experienced insomnia in the first 3 months, they were consulting a gynaecologist throughout pregnancy, etc. were also asked. The data of patients who exercised regularly in the 3 months before pregnancy were assigned a value of 1, and those who did not exercise regularly were assigned a value of 2. Similarly, the data of patients who had good relationships with their parents and spouses and those who did not were assigned values of 1 and 2, respectively. The data of patients who had high levels of trust in close contacts and those who did not were assigned values of 1 and 2, respectively. The data of patients whose per capita family income was greater than 5000 RMB and those of the remaining patients were assigned values of 1 and 2, respectively. The data of patients who had insomnia in the 3 months before pregnancy and those who did not were assigned values of 1 and 2, respectively. The data of patients who had completed senior high school or above and those who had not were assigned values of 1 and 2, respectively. The data of patients who had a job and those who were unemployed were assigned values of 1 and 2, respectively. The data of patients who attained higher than upper secondary school education and those of patients who did not were assigned values of 1 and 2, respectively.

Data on haematological parameters during early pregnancy, such as alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), total bilirubin (TBIL), direct bilirubin (DBIL), indirect bilirubin (IBIL), total bile acid (TBA), cholic acid (CG), total cholesterol (TC), triglyceride (TG), high-density lipoprotein (HDL), and low-density lipoprotein (LDL) levels, were collected.

A cut-off value was calculated on the basis of the specificity and sensitivity of the receiver operating characteristic (ROC) curve. All indicators were dichotomized, with indicators having values greater than the cut-off value being assigned the number 1, whereas those having values less than the cut-off value being assigned the number 2 (refer to Table 2 for specific values).

Table 1 Analysis of variability between training set and internal validation set

Statistical analysis

Statistical analyses were performed via SPSS 26.0 software. Categorical variables were analysed via the χ² test or Fisher’s exact test. Multivariate logistic regression analysis was performed for variables with P < 0.05. Independent risk factors were identified via backwards stepwise regression.

Column‒line plot predictive models of significant predictors from the logistic regression analyses were constructed via the rms package in R software. The total score was calculated by aggregating the scores of each variable to predict the incidence of GDM in pregnant women during pregnancy. A plumb line was drawn at each factor to obtain the corresponding score, and the total score was obtained by adding the scores of all the factors. The predicted probability was obtained by creating a plumb line based on the total score. A greater final score indicated a greater risk.

The model was validated internally and externally, and its discrimination ability and consistency were assessed via ROC curves and the Hosmer–Lemeshow (H–L) test. An area under the curve (AUC) ≤ 0.5 indicated poor discrimination, 0.5 < AUC < 0.7 indicated low discrimination, 0.7 ≤ AUC < 0.9 indicated high discrimination, and an AUC ≥ 0.9 indicated high discrimination. An H–L test with P > 0.05 indicated that the model passed the test, no significant difference existed between the predicted and true values, the model fit well, and the consistency was high. The reverse was true for an H–L test with P < 0.05. The H–L test was conducted with 100 bootstrap samples. The model calibration curve was plotted via R software to evaluate the consistency of the column-line graph prediction model. The test criterion was α = 0.05, and all tests were bilateral.

Results

Characteristics of the study population

From January 2022 to January 2023, the data of 365 pregnant women who had full-term singleton vaginal deliveries were retrospectively collected from the obstetrics department. After excluding the patients whose significant clinical data were missing, data from 137 GDM patients and 128 healthy postpartum women were collected. The obtained data were divided into training and test sets at a 7:3 ratio. Accordingly, 186 samples were included in the training set and 79 samples were included in the test set. Additionally, data from 113 women in labour were collected from Changzhi People’s Hospital during the same period and used as an external validation set.

No significant differences were observed between the training, internal validation, and external validation sets (Tables 1 and 3). In both the training and validation sets, significant differences in weight gain, ALT, and ALP were observed between the healthy and GDM groups (P < 0.05). In the training set, 22 indicators were significantly different between the healthy and GDM groups (P < 0.05). In the internal validation set, 9 indicators differed significantly between the healthy and GDM groups (P < 0.05). In the external validation set, 10 indicators differed significantly between the healthy and GDM groups (P < 0.05, Table 4).

Table 2 ROC curve cut-off

Table 3 Analysis of variability between training set and external validation set

Table 4 Analysis of GDM variability in training set and validation set

Screening for risk factors related to GDM

Binary logistic regression revealed that family history, weight gain, employment or unemployment, income, exercise, insomnia, professional guidance, husband–wife relationship, parental relationship, ALT, AST, ALP, TBIL, DBIL, IBIL, TBA, TC, TG, HDL, and LDL were significantly correlated in the healthy and GDM groups (P < 0.05). According to the multifactorial logistic regression, family history, weight gain, ALT, TBA, TC, and TG were independent risk factors for GDM (Table 5).

Table 5 Logistic analysis of training set and internal validation ensemble set

Construction and validation of the diagnostic model for GDM

The patient data were randomized into training and validation sets at a 7:3 ratio. A diagnostic column-line graph model with 8 predictors was constructed based on the logistic regression results (Fig. 1). The area under the ROC curve (AUROC) of the diagnostic model for the work characteristics of the study subjects was 0.90. The accuracy, sensitivity, specificity, and positive and negative predictive values of the model were 0.83, 0.84, 0.84, 0.76, and 0.90, respectively (Fig. 2; Table 6). The calibration curve displayed good agreement between predicted and actual probabilities (Fig. 3). The decision curve analysis demonstrated clinical decision validity (Figs. 4 and 5). The diagnostic performance of the model was further validated using the validation set.

Fig. 1

Predictive model

Fig. 2

ROC curve of the predictive model and internal test set

Fig. 3

Internal test set calibration curve

Fig. 4

DCA curve of the predictive model

Fig. 5

DCA curve for the internal validation set

Table 6 Confusion matrix analysis (internal text)

External validation

To validate our diagnostic model, the diagnostic efficacy of the model was further validated using an external validation set of 113 pregnant women (43 healthy and 70 with GDM) collected from another centre., and differences between groups were analysed. These pregnant women were selected using the same inclusion and exclusion criteria (Table 5).

Additionally, the model was externally validated using the test set. The calibration curve of the line chart model closely matched the actual values in the test set (Fig. 7). The H − L goodness-of-fit test yielded a P value of 0.975 (Fig. 6). The discriminative ability of the line chart model in predicting the GDM risk was assessed using the AUROC curve, which was 0.72 (Fig. 8; Table 6). The aforementioned validation indicated the good precision and discrimination of the line chart model in predicting GDM risk, and the model’s stability was acceptable.

Fig. 6

ROC curves for the external test set

Fig. 7

External test set calibration curve

Fig. 8

DCA curve for the external validation set

Discussion

This cohort study investigated the correlations between lipids, liver function, and demographics in GDM patients and healthy mothers. Here, we found that ALT, TG, TBA, and TC were independent risk factors for GDM. The logistic regression algorithm screened 8 indicators as predictors, and a prediction model was constructed using ALT, TBA, TC, TG, and whether GDM affects eating habits, the husband–wife relationship, family history, and parental relationships as predictors.

Obesity is known as a major risk factor for GDM [11, 12], which is consistent with the study results. Sleep duration is correlated with glycated haemoglobin. Poorer sleep health, especially lower sleep regularity, is known to predict poorer glucose metabolism in pregnant women, with a corresponding increase in the glycated haemoglobin level and a sharp increase in GDM risk [13]. Our study only preliminarily predicted that insomnia and GDM are associated, and this result is consistent with those of previous studies. According to a prospective study, sleep disorders are a risk factor for GDM, and the optimal amount of sleep for pregnant women is 8–9 h/day [14]. In addition, multitaskers with annual household incomes of more than $30,000 and peak work hours of 32–44 h/week are more likely to develop GDM than single taskers with the same incomes and peak work hours [15]. Our findings are consistent with those of previous studies. No significant differences in growth or developmental indicators were observed in offspring with GDM following individualized nutritional interventions, which could indirectly explain the results of the present study [16, 17].

A prospective combined Mendelian randomization study revealed that the highest quartile of liver function indices (LFIs), including ALT, AST, GGT, ALP, and HSI, was significantly associated with increased GDM risk, and a significant causal relationship existed between ALT and GDM [18]. In another cohort study, women were identified to have the highest risk of GDM based on a combination of parameters, namely, high plasma ferritin levels (˃55.7 ng/mL) and high TG levels (˃1.9 mmol/L). Furthermore, elevated maternal TG concentrations in early pregnancy mediated the relationship between ferritin levels and GDM risk [19]. In the present study, patients with TBIL levels greater than 10.75 µmol/L, DBIL levels greater than 1.95 µmol/L, IBIL levels greater than 7.15 µmol/L, and TBA levels greater than 1.953 µmol/L presented a lower GDM risk. Moreover, AST levels less than 13.65 U/L, ALT levels less than 9.85 U/L, ALP levels less than 92.40 U/L, TC levels less than 6.099 mmol/L, TG levels less than 3.79 mmol/L, HDL levels less than 1.119 mmol/L, and LDL levels less than 2.782 mmol/L were risk factors for GDM. TG, TBA, ALT, and TC were found to be independent risk factors for GDM.

By measuring metabolites produced in organisms through high-throughput technologies such as mass spectrometry and nuclear magnetic resonance, metabolomics has recently provided new insights into disease occurrence and development. Serum metabolism in GDM patients is significantly different from that in healthy pregnant women. Various differentially abundant metabolites and corresponding metabolic pathways, namely, fatty acid metabolism, butyric acid metabolism, bile secretion, and aminoglucose metabolism, have been identified in the serum of GDM patients [20, 21]. A prediction model was constructed that included a family history of diabetes mellitus, a history of GDM in pregnant women, prepregnancy overweight or obesity, a history of hypertension, sedentary time, and high concentrations of monobenzyl phthalate and Q4 monoethyl phthalate as predictors. However, this model has an AUC of 0.827 and is not readily accessible at the clinic; therefore, it has limited value for clinical application. A prediction model for GDM with preeclampsia was constructed by using a combination of cystatin C, uric acid, glutamyltransferase, blood urea nitrogen, and basal systolic blood pressure as predictors; this model had an AUC of 0.8031 [22]. No simple and effective prediction model is currently available for GDM. In our model, the LR algorithm screened 8 indicators as predictors. We then constructed a prediction model by using ALT, TBA, TC, TG, eating habits, the husband–wife relationship, family history, and parental relationships as predictors. Additionally, the DCA results confirmed the positive effect of our model, which confirmed its clinical value.

The findings of this model can aid in early intervention to prevent disease progression. Different predictors (e.g., biochemical indicators and lifestyle) may exert different effects on a patient’s health status. By understanding these effects, doctors can formulate personalized treatment and lifestyle recommendations for their patients to reduce the incidence of the disease or to improve their health status. An analysis of the relationships between these indicators and disease can help researchers explore new biomarkers or interventions to improve their understanding of disease mechanisms, start treatment early, and increase patients’ awareness of their health, thereby motivating patients to adopt positive lifestyle changes. In conclusion, this predictive model is clinically significant because it offers a comprehensive health assessment and intervention strategy based on multiple factors, thereby optimizing treatment and prevention programs for patients. Clinicians can analyse existing clinical guidelines in detail and identify guideline content relevant to the variables predicted by the model. Whether the model can complement or optimize the existing clinical decision-making process by comparing the model variables with the tests or criteria recommended by the guidelines must be evaluated. This study developed a column‒line graph prediction model for GDM based on maternal clinical characteristics and test results from a single centre. The constructed model was internally and externally validated to provide a basis for future studies.

Strengths and limitations

First, our study is a cohort study that can better explain the causal relationship of ALT, ALP, DBIL, TC, TG, and HDL levels with GDM in early pregnancy. Furthermore, we first created a column chart that included biochemical indicators and lifestyle factors to predict GDM. Our external validation results, which were obtained using data from another centre, confirmed that the constructed model had good discriminatory and predictive value. Our study was designed to ensure that the model could be easily used in a clinical setting. We believe that the current model is a simple and valuable tool that can be easily used in clinical practice.

However, our study has limitations. Because of the robustness of our data, we could not monitor the dynamics of ALT, ALP, DBIL, TC, TG, and HDL levels from early to mid-pregnancy. Therefore, the impact of these dynamics on GDM risk could not be explored. This impact could be a topic for future research. Although data from another centre were used for validation, no external validation existed for different populations. Future t extensive external validation studies are recommended to further assess the generalizability and robustness of the model.

Conclusions

In this study, a GDM prediction model that performs well in predicting GDM risk was developed based on routine obstetric tests and information. Using the model, high-risk patients can be easily identified on the basis of biochemical indicators (e.g., ALT, TBA, TC, and TG) and lifestyle factors (e.g., eating habits, the husband–wife relationship, family history, and the parental relationship). In the future, a larger sample size can be used on this basis, and a multicentre prediction model based on a combination of clinical characteristics, examinations, and test results can be constructed and validated.