Early risk stratification of gestational diabetes using interpretable machine learning with first-trimester screening parameters

Abstract

BACKGROUND

Gestational diabetes mellitus (GDM) is usually diagnosed at 24-28 weeks of gestation, when opportunities for early prevention may already be limited. First-trimester maternal characteristics and placental biomarkers may provide earlier risk information, but their combined predictive value remains insufficiently defined.

AIM

To develop and evaluate an interpretable machine learning approach for first-trimester risk stratification of GDM.

METHODS

This retrospective cohort study included singleton pregnancies undergoing first-trimester screening at MacKay Memorial Hospital, a tertiary referral center, between January 2019 and July 2024. Maternal characteristics, obstetric and medical history, mean arterial pressure, ultrasound parameters, and biochemical markers were used to train eight machine learning algorithms. Class imbalance was addressed using synthetic minority oversampling, random oversampling, and random under-sampling. Model performance was evaluated on an independent test set, and interpretability was assessed using SHapley Additive exPlanations (SHAP) and patient-level heatmaps.

RESULTS

Among 2756 singleton pregnancies, 352 women developed GDM (12.8%). Women who developed GDM were older and had higher pregestational body mass index and mean arterial pressure than those without GDM. First-trimester placental biomarkers, including pregnancy-associated plasma protein A, placental growth factor, and free β-human chorionic gonadotropin, were significantly lower in women who subsequently developed GDM. Among all evaluated models and resampling strategies, gradient boosting with random oversampling achieved the best overall performance, with an area under the receiver operating characteristic curve of 0.768, sensitivity of 0.647, specificity of 0.777, positive predictive value of 0.289, negative predictive value of 0.940, and F1 score of 0.400. SHAP analysis identified maternal age, mean arterial pressure, pregestational weight, pregnancy-associated plasma protein A, and placental growth factor as the major contributors to model predictions.

CONCLUSION

An interpretable machine learning approach integrating first-trimester clinical, obstetric, and biochemical parameters may serve as a potential early risk stratification tool for GDM. Given its favorable rule-out performance, the model may help identify women at relatively low risk of GDM during early pregnancy. Further external validation is required before broader clinical implementation.

Key Words: Gestational diabetes mellitus; Machine learning; First trimester; Risk stratification; Pregnancy-associated plasma protein A; Placental growth factor

Core Tip: This retrospective cohort study developed an interpretable machine learning model for early risk stratification of gestational diabetes mellitus using first-trimester clinical and biomarker parameters. In 2756 pregnancies, gradient boosting with random oversampling achieved moderate discrimination with high negative predictive value. Model interpretation using SHapley Additive exPlanations and patient-level heatmaps identified maternal and placental factors as key contributors. This approach may support early identification of at-risk women and facilitate targeted preventive strategies before routine mid-pregnancy screening.

INTRODUCTION

Gestational diabetes mellitus (GDM) is one of the most common metabolic complications of pregnancy and is defined as glucose intolerance with onset or first recognition during gestation[1,2]. Its pathophysiology involves progressive insulin resistance driven by placental hormones together with inadequate β-cell compensation, resulting in maternal hyperglycemia[3]. The prevalence of GDM has increased substantially worldwide in recent decades, with global estimates typically ranging from approximately 7% to 15%, depending on diagnostic criteria and population characteristics[4,5].

GDM is associated with a broad spectrum of adverse maternal and neonatal outcomes. Women with GDM have increased risks of hypertensive disorders of pregnancy, cesarean delivery, and long-term progression to type 2 diabetes mellitus, whereas offspring are at higher risk of macrosomia, neonatal hypoglycemia, and future metabolic disease[6,7]. Large prospective studies have further demonstrated a continuous association between maternal glycemia and adverse pregnancy outcomes, even below traditional diagnostic thresholds[8]. These findings underscore the importance of identifying high-risk women earlier in pregnancy, when preventive strategies may still be clinically meaningful.

Current screening for GDM is typically performed during 24 weeks to 28 weeks of gestation using an oral glucose tolerance test[1]. However, this timing may limit opportunities for early risk modification. Increasing evidence suggests that metabolic and placental alterations related to GDM are already present in the first trimester[9,10]. Accordingly, first-trimester biomarkers, including pregnancy-associated plasma protein A (PAPP-A), placental growth factor (PlGF), and free β-human chorionic gonadotropin (β-hCG), have been investigated for their associations with subsequent GDM, although findings remain heterogeneous across studies[10]. Our previous study showed that lower first-trimester PAPP-A and PlGF levels were associated with subsequent GDM development, and that incorporating these routinely collected biomarkers with maternal clinical characteristics improved predictive performance for early risk assessment[11]. Interventional studies also support the clinical relevance of this approach, as lifestyle intervention among high-risk pregnant women has been shown to reduce the incidence of GDM, and early treatment of GDM diagnosed before mid-pregnancy may improve neonatal outcomes[12,13].

Machine learning (ML) approaches may further improve early prediction by capturing complex and nonlinear relationships among clinical, biochemical, and ultrasound parameters[14,15]. Although numerous ML-based prediction models for GDM have been proposed, many lack independent validation, and their predictive performance may therefore be overestimated[16]. Recent studies have applied various ML approaches for early prediction of GDM using clinical, biochemical, ultrasonographic, and multimodal data (Table 1)[17-28]. However, many previous studies relied on specialized predictors, extensive laboratory testing, or large numbers of variables that may limit implementation in routine prenatal care. In addition, class imbalance handling, patient-level interpretability, and seamless integration into existing first-trimester screening workflows have not been consistently addressed. Therefore, there remains a need for a clinically feasible and interpretable prediction framework based on routinely available first-trimester parameters. While our previous findings supported early GDM risk assessment[11], the potential of ML approaches to improve prediction performance while maintaining clinical interpretability using routinely available first-trimester parameters remains to be established. Therefore, we aimed to develop and evaluate an interpretable ML model using routinely available first-trimester clinical and placental biomarker parameters for early GDM risk stratification.

Table 1 Comparison of previous machine learning studies for gestational diabetes mellitus prediction[17-28].

Ref.	Sample size	Key predictors	Class imbalance	ML algorithms	Best ML performance	ML explainability	External validation	Calibration analysis
Wu et al[17]	31811	Clinical, biochemical, lipid, thyroid, and obstetric variables; 73-variable model and simplified 7-variable LR model	Not reported	DNN, SVM, KNN, and LR	DNN using 73 variables, AUC: 80%; 7-variable LR, AUC: 77%	Not reported	No	No
Xiong et al[18]	490	Routine blood tests, hepatic and renal function markers, and coagulation markers, particularly PT and aPTT	Not reported	SVM and LightGBM	SVM using PT and aPTT, sensitivity: 88.3%, specificity: 99.47%, and AUC: 94.2%	Not reported	No	No
Kaya et al[19]	97	First-visit venous plasma glucose level, maternal BMI, family history of DM, smoking, and obstetric history	Not reported	Extra trees, average blender, LightGBM, XGBoost, LR, and RF	XGBoost, AUC: 55.0%, accuracy: 66.7%, sensitivity: 80.0%, and specificity: 50.0% in nulliparous women; AUC: 73.3%, accuracy: 72.7%, sensitivity: 40.0%, and specificity: 100.0% in primiparous women	SHAP	No	No
Li et al[20]	7594	Forty-five first-trimester features; top predictors included pre-pregnancy BMI and maternal abdominal circumference at pregnancy initiation, and FPG and HbA1c at the end of the first trimester	Not reported	LR, XGBoost, RF, and other ML algorithms	XGBoost, AUC: 75% at pregnancy initiation and 99% at the end of the first trimester in the XHCM cohort; external validation AUC: 83% in the SPNPH cohort	Feature importance analysis	Yes	No
Zorlu et al[21]	400	Maternal characteristics, BMI, PAPP-A, and free β-hCG	Not reported	RF, GradBoost, and LR	GradBoost, AUC: 71.5% and accuracy: 71.3%	Not reported	No	No
Ni et al[22]	956	Common first-trimester clinical and laboratory variables selected by Spearman correlation analysis and Boruta algorithm, including pre-pregnancy BMI, SBP, and HDL-C	Not reported	LR, RF, XGBoost, LightGBM, MLP, KNN, and SVM	LR, AUC: 78.7% (95%CI: 72.3%-85.0%); RF, AUC: 77.6% (95%CI: 71.1%-84.1%)	Not reported	No	Yes
Zaky et al[23]	138	History of high glucose/diabetes, insulin, HOMA-IR, uric acid, cholesterol, urea, PT, NT-proBNP, thyroid markers, and routine blood markers	Not reported	RF, GradBoost, AdaBoost, DT, LR, SVM, Gaussian NB, KNN, CatBoost, XGBoost, LightGBM, and stacking ensemble	Stacking ensemble, accuracy: 88.8%, precision: 87.3%, sensitivity: 92.1%, and F1 score: 89.6%	SHAP	No	No
Bigdeli et al[24]	106	Age, BMI, previous abortion history, FPG, demographic variables, medical history, and clinical findings	SMOTE	DT, MLP, KNN, NB, RF, and XGBoost	RF, accuracy: 89%, precision: 86%, sensitivity: 92%, and AUC: 94%	Not reported	No	No
Pazaras et al[25]	797	Maternal demographics, obstetric history, lifestyle factors, and FFQ-derived dietary micronutrient intake	SMOTE, borderline SMOTE, adaptive synthetic sampling, and SMOTE-Tomek	LR, RF, extra trees, XGBoost, LightGBM, CatBoost, GradBoost, AdaBoost, and MLP	LR without resampling, AUC: 66.4% (95%CI: 54.2%-77.7%), sensitivity: 78.3%, and NPV: 93.2%; reduced 9-feature model, AUC: 71.2% (95%CI: 58.9%-82.5%)	SHAP	No	Yes
Prashanthan and Prashanthan[26]	10000	Synthetic demographic characteristics, clinical risk factors, and first-trimester laboratory parameters including random blood sugar, post-prandial blood sugar, HbA1c, and OGTT values	SMOTE	Multiple ML algorithms	Best model, accuracy: 71.7% and AUC: 76.9%	SHAP	No	No
Zhai et al[27]	534	BMI, SAT, VAT, maternal clinical characteristics, and first-trimester ultrasonographic indicators	IPW	XGBoost, ANN, SVM, MLR, and RF	XGBoost with GA-selected features, internal test AUC: 96.2%; external validation AUC: 87.8%, sensitivity: 70.0%, and specificity: 93.5%	GA-based feature selection and heatmaps	Yes	No
Louzoun et al[28]	596 twin pregnancies	WBC count, platelet levels, BMI, and previous GDM	Not reported	LightGBM, XGBoost, and LR	LightGBM, AUC: 72% (95%CI: 69%-75%); detection rates: 28% and 42% at false-positive rates of 10% and 20%, respectively	Not reported	No	No
Present study	2756	Maternal age, MAP, pregestational weight, PAPP-A, and PlGF	SMOTE, ROS, and RUS	LR, GradBoost, XGBoost, AdaBoost, LightGBM, SVM, RF, and MLP	GradBoost + ROS, AUC: 76.8%, sensitivity: 64.7%, specificity: 77.7%, PPV: 28.9%, NPV: 94.0%, and F1 score: 40.0%	SHAP and patient-level heatmaps	No	Yes

ML: Machine learning; LR: Logistic regression; PT: Prothrombin time; aPTT: Activated partial thromboplastin time; BMI: Body mass index; DM: Diabetes mellitus; FPG: Fasting plasma glucose; HbA1c: Glycated hemoglobin; PAPP-A: Pregnancy-associated plasma protein A; β-hCG: β-human chorionic gonadotropin; SBP: Systolic blood pressure; HDL-C: High-density lipoprotein cholesterol; HOMA-IR: Homeostatic model assessment for insulin resistance; NT-proBNP: N-terminal pro-B-type natriuretic peptide; FFQ: Food frequency questionnaire; OGTT: Oral glucose tolerance test; SAT: Subcutaneous adipose tissue; VAT: Visceral adipose tissue; WBC: White blood cell count; GDM: Gestational diabetes mellitus; MAP: Mean arterial pressure; PlGF: Placental growth factor; IPW: Inverse probability weighting; SMOTE: Synthetic minority oversampling technique; RUS: Random undersampling; ROS: Random oversampling; DNN: Deep neural network; SVM: Support vector machine; KNN: K-nearest neighbors; LightGBM: Light gradient boosting machine; XGBoost: Extreme gradient boosting; RF: Random forest; GradBoost: Gradient boosting; MLP: Multilayer perceptron; AdaBoost: Adaptive boosting; DT: Decision tree; Gaussian NB: Gaussian naïve Bayes; CatBoost: Categorical boosting; NB: Naïve Bayes; ANN: Artificial neural network; MLR: Multivariate logistic regression; CI: Confidence interval; AUC: Area under the receiver operating characteristic curve; NPV: Negative predictive value; GA: Genetic algorithm; SHAP: SHapley Additive exPlanations; PPV: Positive predictive value.

MATERIALS AND METHODS

Study design and population

This retrospective cohort study included singleton pregnant women who underwent routine first-trimester screening at MacKay Memorial Hospital, a tertiary referral center, between January 2019 and July 2024. All participants underwent GDM screening at 24-28 weeks of gestation using either a 75-g or 100-g oral glucose tolerance test. GDM was diagnosed according to either the one-step International Association of Diabetes and Pregnancy Study Groups (IADPSG) criteria or the two-step Carpenter-Coustan or National Diabetes Data Group (NDDG) criteria, according to the clinical protocol used during the study period. Women meeting any of these diagnostic criteria were classified as having GDM for subsequent analyses.

Women with multiple pregnancy, pregestational diabetes mellitus, preeclampsia, major fetal anomalies, autoimmune disease, chronic renal disease, chronic hepatic disease, or maternal age < 18 years or > 50 years were excluded, whereas women with chronic hypertension or cardiovascular disease were retained because these conditions are clinically relevant to GDM risk stratification. This study was approved by the Institutional Review Board of MacKay Memorial Hospital (No. 26MMHIS097e).

Clinical, biochemical, and ultrasound feature collection

Clinical and demographic data were extracted from electronic medical records, including maternal age, pregestational body mass index, previous GDM, previous gestational hypertension, previous preeclampsia, family history of diabetes mellitus, previous macrosomia (fetal birth weight ≥ 4000 g), history of polycystic ovary syndrome (PCOS), conception via in vitro fertilization (IVF), chronic hypertension, and cardiovascular disease. Family history of diabetes mellitus was defined as chronic diabetes in a first- or second-degree relative.

First-trimester parameters included gestational age at scan, crown-rump length (CRL), nuchal translucency (NT), mean arterial pressure (MAP), uterine artery pulsatility index (PI), and biochemical markers, including PAPP-A, PlGF, and free β-hCG. Biochemical markers were measured using a Kryptor analyzer (Brahms GmbH, Germany). Ultrasound examinations were performed using a Voluson E10 ultrasound system (GE Healthcare, Zipf, Austria) equipped with a 3-9 MHz transabdominal probe. CRL, NT thickness, and uterine artery PI were assessed in accordance with the guidelines of the Fetal Medicine Foundation.

Model development and validation

The dataset was randomly divided into a development set (90%) and an independent test set (10%) using stratified sampling according to GDM outcome to preserve the original class distribution. Before model training, missing continuous variables were imputed using mean values, whereas categorical variables were imputed using the most frequent category. The proportion of missing data for each variable is presented in Supplementary Table 1. Model training and internal validation were performed within the development set using five-fold stratified cross-validation. To address class imbalance while minimizing data leakage, synthetic minority oversampling technique (SMOTE), random oversampling (ROS), and random undersampling (RUS) were applied exclusively to the training folds and were not applied to the validation folds or independent test set.

Eight ML algorithms were evaluated, including logistic regression, gradient boosting (GradBoost), extreme GradBoost, adaptive boosting, light GradBoost machine, support vector machine, random forest, and multilayer perceptron. Continuous predictors were retained as numeric variables, whereas clinical history variables were coded as binary indicators. For logistic regression, support vector machine, and multilayer perceptron models, z-score standardization was applied after imputation, whereas the remaining algorithms were fitted without feature scaling. Hyperparameter optimization was not performed, and default model settings with fixed random seeds were used to improve reproducibility. Nested or repeated cross-validation was not adopted because an independent hold-out test set was prioritized to assess model generalizability. After cross-validation, each final model was retrained using the entire development set and subsequently evaluated on the independent test set.

The primary performance metric was the area under the receiver operating characteristic curve (AUC-ROC). Secondary performance metrics included sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, and accuracy. Binary classification metrics were calculated using a fixed probability threshold of 0.50. The final best-performing model was selected according to overall test-set discrimination performance, primarily based on AUC-ROC and F1 score. Ninety-five percent confidence intervals were calculated for the AUC-ROC estimates to improve interpretation of model stability. As a supplementary analysis, calibration curves and Brier scores were generated for the best-performing models under the ROS, RUS, and SMOTE strategies.

Model interpretability analysis

Model interpretability was evaluated for the final best-performing model using SHapley Additive exPlanations (SHAP), which decompose model predictions into additive feature-level contributions. SHAP values were calculated on the independent test set. The mean absolute SHAP value was used to quantify global feature importance, representing the average magnitude of each feature’s contribution across all predictions. The direction of feature contribution was determined according to the sign of the SHAP values. SHAP summary plots were generated to illustrate global feature importance and overall contribution patterns, whereas patient-level SHAP heatmaps were used to visualize individualized feature contributions.

Statistical analysis

Statistical analyses were performed using IBM SPSS Statistics for Windows, version 29.0 (IBM Corp., Armonk, NY, United States), and ML analyses were conducted using Python version 3.12.4. The Kolmogorov-Smirnov test was used to assess the normality of continuous variables. Continuous variables were compared using the Student’s t-test or Mann-Whitney U test, whereas categorical variables were analyzed using the χ² test or Fisher’s exact test depending on expected cell counts. Continuous variables are presented as mean ± SD or median with interquartile range according to data distribution, whereas categorical variables are presented as n (%). All statistical tests were two-tailed, and a P value < 0.05 was considered statistically significant.

RESULTS

Baseline maternal characteristics

A total of 2756 singleton pregnancies were included in the final analysis, of which 352 women (12.8%) developed GDM and 2404 did not (Figure 1). Among the 352 women with GDM, 195 (55.4%) were diagnosed according to the IADPSG criteria, 42 (11.9%) according to the Carpenter-Coustan criteria, and 115 (32.7%) according to the NDDG criteria. Baseline maternal characteristics are summarized in Table 2. Women who developed GDM were significantly older than those without GDM (34.0 years vs 33.0 years, P < 0.001) and had higher pregestational BMI (22.9 kg/m² vs 21.5 kg/m², P < 0.001). Previous GDM (8.5% vs 0.8%, P < 0.001), previous macrosomia (1.7% vs 0.3%, P = 0.004), conception via IVF (13.6% vs 9.2%, P = 0.013), chronic hypertension (3.1% vs 0.7%, P < 0.001), and cardiovascular disease (1.1% vs 0.04%, P = 0.001) were also significantly more common in the GDM group. No significant differences were observed for previous gestational hypertension, previous preeclampsia, family history of diabetes mellitus, or PCOS history.

Figure 1 Flowchart of participant selection and study population enrollment for gestational diabetes mellitus prediction analysis. GDM: Gestational diabetes mellitus.

Table 2 Baseline characteristics of the study population, n (%)/median (interquartile range).

Variables	GDM (n = 352)	Non-GDM (n = 2404)	P value
Age (years)	34.0 (31.0-37.0)	33.0 (30.0-35.0)	< 0.001a
Pregestational BMI (kg/m2)	22.9 (20.7-27.1)	21.5 (19.6-23.9)	< 0.001a
Previous GDM	30 (8.5)	19 (0.8)	< 0.001a
Previous gestational hypertension	2 (0.6)	11 (0.5)	0.671
Previous preeclampsia	5 (1.4)	17 (0.7)	0.181
Family history of diabetes mellitus	30 (8.5)	182 (7.6)	0.561
Previous macrosomia	6 (1.7)	7 (0.3)	0.004a
PCOS history	0 (0)	9 (0.4)	0.613
IVF	48 (13.6)	221 (9.2)	0.013a
Chronic hypertension	11 (3.1)	17 (0.7)	< 0.001a
Cardiovascular disease	4 (1.1)	1 (< 0.1)	0.001a

^aP < 0.05.

GDM: Gestational diabetes mellitus; BMI: Body mass index; PCOS: Polycystic ovary syndrome; IVF: In vitro fertilization.

First-trimester clinical, ultrasound, and biochemical parameters

First-trimester clinical, ultrasound, and biochemical parameters are presented in Table 3. Women who subsequently developed GDM had significantly higher MAP values than women without GDM (84.7 mmHg vs 81.7 mmHg, P < 0.001). First-trimester placental biomarkers were significantly lower in the GDM group, including PAPP-A (5.07 IU/L vs 5.80 IU/L, P < 0.001), PlGF (41.80 pg/mL vs 43.00 pg/mL, P = 0.049), and free β-hCG (38.60 IU/L vs 41.40 IU/L, P = 0.020). No significant differences were observed in gestational age at scan, CRL, NT thickness, or uterine artery PI between the two groups.

Table 3 First-trimester clinical, ultrasound, and biochemical parameters of the study population, median (interquartile range).

Variables	GDM (n = 352)	Non-GDM (n = 2404)	P value
GA at scan (weeks)	12.7 (12.4-13.0)	12.7 (12.4-13.0)	0.277
CRL (mm)	66.6 (62.4-70.5)	66.6 (62.9-70.4)	0.385
NT (mm)	1.8 (1.7-2.0)	1.8 (1.7-2.0)	0.879
MAP (mmHg)	84.7 (77.3-93.0)	81.7 (75.7-88.0)	< 0.001a
PAPP-A (IU/L)	5.07 (3.26-6.96)	5.80 (4.04-8.11)	< 0.001a
PlGF (pg/mL)	41.80 (29.66-53.26)	43.00 (31.31-57.00)	0.049a
Free β-hCG (IU/L)	38.60 (25.80-58.80)	41.40 (28.60-61.00)	0.020a
Uterine artery PI	1.60 (1.32-1.97)	1.58 (1.31-1.89)	0.072

^aP < 0.05.

GDM: Gestational diabetes mellitus; GA: Gestational age; CRL: Crown-rump length; NT: Nuchal translucency; MAP: Mean arterial pressure; PAPP-A: Pregnancy-associated plasma protein A; PlGF: Placental growth factor; β-hCG: β-human chorionic gonadotropin; PI: Pulsatility index.

Performance of ML models

The predictive performance of the ML models under different class imbalance handling strategies is summarized in Table 4 and Figure 2. Model performance varied substantially according to the resampling strategy used. Models trained on the original imbalanced dataset generally demonstrated high specificity but low sensitivity, indicating limited ability to identify women who subsequently developed GDM. In contrast, resampling strategies improved sensitivity across multiple models, although this was often accompanied by reduced specificity. Overall, GradBoost with ROS achieved the best overall discrimination and balanced classification performance on the independent test set, with an AUC-ROC of 0.768, sensitivity of 0.647, specificity of 0.777, PPV of 0.289, NPV of 0.940, F1 score of 0.400, and accuracy of 0.761. Calibration curves and Brier scores for the best-performing models under the ROS, RUS, and SMOTE strategies are presented in Supplementary Figure 1.

Figure 2 Receiver operating characteristic curves comparing the predictive performance of multiple machine learning models for gestational diabetes mellitus prediction under different class imbalance handling strategies on the independent test set. A: Original dataset without resampling; B: Synthetic minority oversampling technique; C: Random undersampling; D: Random oversampling. The evaluated models included logistic regression, extreme gradient boosting, adaptive boosting, gradient boosting, light gradient boosting machine, support vector machine, random forest, and multilayer perceptron. The X-axis represents the false positive rate, and the Y-axis represents the true positive rate. The diagonal dashed line indicates random classification performance. The area under the curve values for each model are shown in the corresponding legends. SMOTE: Synthetic minority oversampling technique; RUS: Random undersampling; ROS: Random oversampling; LogReg: Logistic regression; XGBoost: Extreme gradient boosting; AdaBoost: Adaptive boosting; GradBoost: Gradient boosting; LightGBM: Light gradient boosting machine; SVM: Support vector machine; RandForest: Random forest; MLP: Multilayer perceptron; AUC: Area under the curve.

Table 4 Performance metrics of different machine learning models.

Sampler	Model	AUC-ROC	Sensitivity	Specificity	PPV	NPV	F1 score	Accuracy
Original	LogReg	0.676 (0.565-0.784)	0.147	0.992	0.714	0.892	0.244	0.888
	XGBoost	0.720 (0.618-0.808)	0.118	0.988	0.571	0.888	0.195	0.880
	AdaBoost	0.674 (0.548-0.792)	0.147	0.975	0.455	0.891	0.222	0.873
	GradBoost	0.731 (0.621-0.832)	0.147	0.983	0.556	0.891	0.233	0.880
	LightGBM	0.702 (0.600-0.795)	0.118	0.988	0.571	0.888	0.195	0.880
	SVM	0.703 (0.592-0.808)	0.088	1.000	1.000	0.886	0.162	0.888
	RandForest	0.713 (0.605-0.813)	0.029	0.996	0.500	0.880	0.056	0.877
	MLP	0.639 (0.531-0.740)	0.147	0.938	0.250	0.887	0.185	0.841
SMOTE	LogReg	0.675 (0.562-0.787)	0.559	0.591	0.161	0.905	0.250	0.587
	XGBoost	0.694 (0.588-0.795)	0.235	0.979	0.615	0.901	0.340	0.888
	AdaBoost	0.674 (0.564-0.783)	0.353	0.901	0.333	0.908	0.343	0.833
	GradBoost	0.659 (0.553-0.762)	0.206	0.971	0.500	0.897	0.292	0.877
	LightGBM	0.696 (0.597-0.783)	0.059	0.979	0.286	0.881	0.098	0.866
	SVM	0.675 (0.564-0.777)	0.618	0.711	0.231	0.930	0.336	0.699
	RandForest	0.708 (0.611-0.803)	0.147	0.983	0.556	0.891	0.233	0.880
	MLP	0.632 (0.531-0.736)	0.265	0.905	0.281	0.898	0.273	0.826
ROS	LogReg	0.677 (0.565-0.789)	0.588	0.587	0.167	0.910	0.260	0.587
	XGBoost	0.701 (0.605-0.790)	0.176	0.926	0.250	0.889	0.207	0.833
	AdaBoost	0.715 (0.602-0.819)	0.588	0.752	0.250	0.929	0.351	0.732
	GradBoost	0.768 (0.679-0.849)	0.647	0.777	0.289	0.940	0.400	0.761
	LightGBM	0.720 (0.620-0.811)	0.206	0.975	0.538	0.897	0.298	0.880
	SVM	0.707 (0.591-0.813)	0.618	0.723	0.239	0.931	0.344	0.710
	RandForest	0.724 (0.618-0.823)	0.088	0.996	0.750	0.886	0.158	0.884
	MLP	0.610 (0.504-0.723)	0.235	0.909	0.267	0.894	0.250	0.826
RUS	LogReg	0.680 (0.567-0.789)	0.559	0.624	0.173	0.910	0.264	0.616
	XGBoost	0.736 (0.642-0.821)	0.706	0.640	0.216	0.939	0.331	0.649
	AdaBoost	0.662 (0.565-0.761)	0.647	0.612	0.190	0.925	0.293	0.616
	GradBoost	0.717 (0.614-0.816)	0.706	0.628	0.211	0.938	0.324	0.638
	LightGBM	0.737 (0.629-0.826)	0.735	0.570	0.194	0.939	0.307	0.591
	SVM	0.708 (0.594-0.805)	0.706	0.607	0.202	0.936	0.314	0.620
	RandForest	0.740 (0.640-0.828)	0.706	0.616	0.205	0.937	0.318	0.627
	MLP	0.643 (0.524-0.753)	0.676	0.525	0.167	0.920	0.267	0.543

AUC-ROC: Area under the receiver operating characteristic curve; NPV: Negative predictive value; PPV: Positive predictive value; SMOTE: Synthetic minority oversampling technique; RUS: Random undersampling; ROS: Random oversampling; LogReg: Logistic regression; XGBoost: Extreme gradient boosting; AdaBoost: Adaptive boosting; GradBoost: Gradient boosting; LightGBM: Light gradient boosting machine; SVM: Support vector machine; RandForest: Random forest; MLP: Multilayer perceptron.

Model interpretability analysis

Model interpretability analysis of the final GradBoost model is presented in Figure 3. Figure 3A shows the SHAP summary plot, which identified maternal age, MAP, pregestational weight, PAPP-A, and PlGF as major contributors to model predictions. Higher maternal age, higher MAP, and higher pregestational weight were generally associated with increased predicted GDM risk, whereas higher PAPP-A and PlGF levels tended to contribute to lower predicted risk. Figure 3B presents patient-level SHAP heatmaps illustrating individualized prediction patterns and variable-specific contribution distributions across participants, supporting model interpretability at both global and individual levels.

Figure 3 Model interpretability analysis of the gradient boosting model using SHapley Additive exPlanations. A: SHapley Additive exPlanations summary plot demonstrating the relative contribution and impact of clinical, biochemical, and ultrasound variables on gestational diabetes mellitus prediction; B: Heatmap of SHapley Additive exPlanations values across individual participants, illustrating feature importance patterns and model output distribution. Warmer colors indicate positive contributions to the predicted risk of gestational diabetes mellitus, whereas cooler colors indicate negative contributions. Greater color intensity reflects larger contribution magnitudes. SHAP: SHapley Additive exPlanations; MAP: Mean arterial pressure; BW: Weight; PAPP-A: Pregnancy-associated plasma protein A; PlGF: Placental growth factor; GDM: Gestational diabetes mellitus; BH: Body height; NT: Nuchal translucency; PI: Pulsatility index; CRL: Crown-rump length; β-hCG: β-human chorionic gonadotropin; PCOS: Polycystic ovary syndrome; IVF: In vitro fertilization; DM: Diabetes mellitus.

DISCUSSION

In this retrospective cohort study, we developed and evaluated a prediction model for early GDM risk stratification using routinely available first-trimester maternal characteristics, ultrasound parameters, and placental biomarkers. Among the evaluated models and resampling strategies, GradBoost with ROS achieved the best overall discrimination and balanced classification performance, with an AUC-ROC of 0.768 and a high NPV of 0.940. These results indicate that routinely available first-trimester parameters may help identify women at relatively low risk of subsequent GDM before routine mid-pregnancy screening[1,2]. Because first-trimester screening is already integrated into routine prenatal care, the same clinical visit may also provide an opportunity for earlier metabolic risk assessment before the conventional timing of GDM diagnosis. Maternal age, MAP, pregestational weight, PAPP-A, and PlGF were identified as major contributors to risk prediction. These observations are consistent with previous studies reporting associations between first-trimester placental biomarkers, maternal metabolic characteristics, and subsequent GDM development[10,11].

PAPP-A is involved in the regulation of insulin-like growth factor bioavailability and maternal glucose metabolism, and lower first-trimester PAPP-A levels have been associated with insulin resistance and later development of GDM. PlGF also plays an important role in placental angiogenesis and metabolic adaptation during pregnancy, and abnormal placental vascular function may contribute to impaired glucose homeostasis and beta-cell dysfunction[10,11]. These findings support the concept that GDM is not solely a disorder of maternal glucose metabolism, but may also reflect early abnormalities in placental development and maternal vascular adaptation. Elevated MAP during early pregnancy may likewise reflect underlying insulin resistance and vascular-metabolic dysfunction, and recent studies have demonstrated associations between higher early-pregnancy blood pressure patterns and increased GDM risk[29,30]. The integration of maternal metabolic, hemodynamic, and placental biomarkers within the model therefore appears biologically reasonable. Maternal and placental factors also remained among the most influential variables in the feature importance analysis, supporting the clinical relevance of the prediction results.

Our findings further extend those of our previous study, in which conventional logistic regression analysis demonstrated associations between first-trimester maternal characteristics, placental biomarkers, and subsequent GDM development in Chinese women[11]. In the current study, we included a larger cohort and evaluated multiple prediction models with different class imbalance handling strategies. Compared with the previous conventional statistical model, the current models showed improved overall discrimination and more balanced classification performance after resampling procedures were applied. Conventional regression models are primarily designed to estimate independent associations between variables, but may be less flexible in capturing nonlinear or combined relationships among maternal characteristics, hemodynamic parameters, and placental biomarkers during early pregnancy. Although the improvement in discrimination observed with the GradBoost model was modest, the potential value of ML may extend beyond overall predictive performance by enabling the identification of complex patterns that are less readily recognized using conventional statistical models. Our findings therefore should be viewed as an extension of a previously established maternal and placental biomarker framework rather than a replacement for conventional statistical modeling.

Several previous studies have explored the use of ML methods for early GDM prediction; however, study populations, predictor selection, sample size, and reported model performance have varied substantially across studies[16,17]. Some studies have reported favorable predictive performance using extensive biochemical or clinical variables; however, application of these models in routine prenatal care may be limited by variable availability, data complexity, and differences in local screening workflows[17,20]. Other studies have used relatively small cohorts or specialized predictors, which may limit generalizability when applied to broader obstetric populations[19,24]. However, class imbalance handling and patient-level interpretability have not been consistently incorporated into previous model development[16]. These issues are particularly important in GDM prediction because the outcome is inherently imbalanced, and clinically useful prediction models should ideally provide interpretable information in addition to risk estimation. In this study, we focused on routinely available first-trimester maternal characteristics, ultrasound parameters, and standard prenatal screening biomarkers commonly obtained in clinical practice. The use of clinically accessible variables together with patient-level risk interpretation may facilitate integration into routine early pregnancy care.

Class imbalance handling had a substantial influence on model performance. In the original dataset, the models generally demonstrated high specificity but relatively limited sensitivity for identifying women who subsequently developed GDM. After resampling procedures were applied, classification performance became more balanced, particularly in the ROS-based models. This finding is important because GDM prediction datasets are inherently imbalanced, with non-GDM pregnancies substantially outnumbering GDM cases in most obstetric populations[16,22]. Our cohort also reflected the prevalence distribution encountered in routine clinical practice. Under these conditions, model evaluation based primarily on overall accuracy or specificity may overestimate predictive performance while failing to adequately identify women at increased metabolic risk. The present results also suggest that assessment of early GDM prediction models should not rely solely on AUC, particularly in clinically imbalanced datasets. Resampling procedures improved sensitivity and F1 score while maintaining acceptable overall discrimination, indicating that imbalance handling strategies may improve the clinical utility of early GDM prediction models. Among the evaluated approaches, ROS showed more stable overall performance than SMOTE and RUS in our cohort. Although ROS may theoretically increase the risk of overfitting because of duplicated minority class observations, this concern was not apparent in the independent test-set results. One possible explanation is that ROS preserved the original clinical data structure while increasing representation of GDM cases, whereas excessive undersampling may reduce available clinical information and synthetic resampling methods may be more sensitive to variability in heterogeneous obstetric datasets[16].

Interpretability remains an important consideration when prediction models are applied in pregnancy care. Feature importance analysis identified maternal age, MAP, pregestational weight, PAPP-A, and PlGF as major contributors to model predictions, findings that were broadly consistent with established knowledge regarding GDM risk factors[2,3,10]. Patient-level SHAP analysis further illustrated how the relative contributions of individual variables varied across women, providing a more transparent interpretation of model predictions. The importance of interpretable prediction frameworks has also been increasingly emphasized in healthcare applications, particularly when models are intended to support clinical decision-making rather than purely automated systems[31,32].

The present findings may have several implications for early pregnancy care. Because the variables included in the model are routinely obtained during first-trimester prenatal screening, early GDM risk stratification may be performed without substantial additional testing or major modification of existing clinical workflows. Earlier identification of women who may be at increased risk of GDM could potentially facilitate consideration of preventive strategies evaluated in previous studies[12,13]. The relatively high NPV suggests potential utility in identifying women at relatively low risk of GDM. The relatively low PPV may also lead to false-positive classification and unnecessary surveillance during early pregnancy. The model should therefore be regarded as a possible early risk stratification tool rather than a definitive clinical prediction tool. Conventional mid-pregnancy glucose screening remains essential for diagnosis.

This study has several strengths. First, we included a relatively large cohort of singleton pregnancies with comprehensive first-trimester maternal characteristics, ultrasound parameters, and placental biomarker data obtained during routine prenatal screening, supporting potential clinical feasibility and practical applicability without substantial additional testing. Second, multiple prediction models and different resampling strategies were systematically evaluated using an independent test set, allowing comprehensive comparison of model performance under class imbalance conditions. Third, patient-level feature importance analysis was incorporated to improve interpretability of the prediction results and to provide individualized information regarding the relative contribution of maternal and placental factors to model prediction.

Several limitations should also be acknowledged. First, this was a retrospective single-center study, which may limit generalizability to other ethnic populations or healthcare systems. Differences in ethnicity, prenatal screening protocols, and biomarker assay platforms across healthcare settings may influence model calibration and transportability. Second, external validation was not performed, and the robustness of the model in independent populations remains to be established. Third, some potentially relevant metabolic variables, including fasting glucose, glycated hemoglobin, and longitudinal glycemic measurements during pregnancy, were not included in the analysis. Fourth, different GDM diagnostic criteria were used during the study period, which may have introduced some degree of outcome heterogeneity, and criterion-specific sensitivity analyses were not performed because of the limited number of cases in certain diagnostic subgroups. Future multicenter prospective studies with external validation and standardized diagnostic criteria are warranted to further evaluate the robustness and transportability of the proposed model.

CONCLUSION

Our interpretable ML model integrating first-trimester maternal characteristics, ultrasound parameters, and placental biomarkers may serve as a possible early risk stratification tool for GDM and demonstrated favorable rule-out performance during early pregnancy. The model should be regarded as an adjunctive approach to complement, rather than replace, established screening strategies. Further multicenter prospective studies and external validation are required before broader clinical application can be considered.