Deep learning model enhances gestational diabetes prediction accuracy

Deep learning model enhances gestational diabetes prediction accuracy | Image Credit: © Chinnapong - © Chinnapong - stock.adobe.com.

A risk prediction model based on multi-layer perceptron (MLP) may improve gestational diabetes mellitus (GDM) prediction, according to a recent study published in Gynecological Endocrinology.¹

GDM occurs in approximately 5% to 30% of pregnant patients, shifting toward a younger population over time. Regional and ethnic variation has also been reported for GDM,² and the condition can lead to adverse outcomes in both mothers and infants.¹ This highlights GDM as a significant public health concern.

“Due to its adeptness in handling intricate interactions and nonlinearities among input variables, [deep learning (DL)] has demonstrated superiority over traditional regression models in certain studies,” wrote investigators. “However, the utility of DL when utilizing a limited set of variables in the context of GDM remains uncertain in comparison to logistic regression.”

To compare novel DL techniques to traditional logistic regression, investigators conducted a retrospective cohort study. Participants included individuals with complete relevant prenatal health care data delivering from 2008 to 2018.

Folic acid supplementation was reported, defined by taking 0.4 mg of folic acid per day starting 3 months before pregnancy preparation. Exclusion criteria included pregestational diabetes mellitus, missing or indistinct prenatal health care data, and abnormal values in demographic characteristics or laboratory biomarkers.

The first stage of the analysis involved determining high-risk factors of GDM, while the second handling imbalance data. Finally, the third stage involved constructing and evaluating a classification model.

Multivariate logistic regression and nomogram model were employed to establish a baseline prediction model for GDM. Afterward, investigators developed and trained a neural network model for early GDM risk prediction. This allowed comprehensive assessment of risk capabilities.

Three layers were included in the neural network GDM predictor. The first layer was a linear layer followed by a nonlinear rectified linear unit activation function. The second layer was also linear and included 64 hidden neurons. Finally, an output layer with 2 neurons was included, generating a real number ranging from 0 to 1.

The cohort encompassed 32 variables selected from 42. Of these, 7 variables significantly differed between pregnant women with GDM and those without GDM. These included history of hypertension, family history of hypertension, family history of diabetes mellitus, folic acid supplementation, age, age at menarche, education level, and prepregnancy body mass index (BMI).

GDM patients more often took folic acid supplementation, had a lower education level, were older, had a lower age at menarche, and had a higher average BMI compared to the non-GDM group. Hematological, renal, and liver function indicators significantly differed between groups.

Higher HGB, WBC, and PLT levels were reported in the GDM group, at 126.01 g/L, 8.94 × 10⁹/L, and 241.78 × 10⁹/L, respectively, vs 123.20 g/L, 8.21 × 10⁹/L, and 227.22 × 10⁹/L, respectively, in the non GDM group. For indicators of renal function, women with GDM had lower Scr levels but higher BUN levels.

When evaluating liver function, the GDM group had higher ALT and AST levels than the non-GDM group, at 19.54 U/L and 18.29 U/L, respectively, vs 16.96 U/L and 17.74 U/L, respectively. TBIL and DBIL levels were lower at 10.35 μmol/L and 3.50 μmol/L, respectively, for GDM vs 10.60 μmol/L and 3.63 μmol/L, respectively, for no GDM.

The model was assessed based on the area under the receiver operating characteristic curve (auROC), average precision (auPR) and F1 score. An auROC of 0.943, auPR of 0.855, and F1 score of 0.879 were reported in validation set, vs 0.774, 0.272, and 0.377, respectively, for the MLP model.

“The inclusion of genetic testing data would further enhance research into susceptible genes and their pathological mechanisms related to GDM, enabling the provision of personalized preventive and therapeutic strategies based on individual susceptibility information,” concluded investigators.

References

1. Zhao M, Su X, Huang L. Early gestational diabetes mellitus risk predictor using neural network with NearMiss. Gynecol Endocrinol. 2025;41(1):2470317. doi:10.1080/09513590.2025.2470317
2. Cho NH, Shaw JE, Karuranga S, et al. IDF Diabetes Atlas: Global estimates of diabetes prevalence for 2017 and projections for 2045. Diabetes Res Clin Pract. 2018;138:271-281. doi:10.1016/j.diabres.2018.02.023