Gestational diabetes mellitus (GDM) and hypertensive disorders of pregnancy (HDP) are two common pregnancy-related complications that independently increase the risk of adverse maternal and neonatal outcomes, including preterm birth, placental abruption, fetal growth restriction, and perinatal mortality (1). Recent epidemiological evidence suggests that the prevalence of GDM has risen to approximately 14% (2), whereas the prevalence of HDP has increased to around 10% (3). Notably, the incidence of both GDM and HDP has been rising in recent years, with certain studies indicating that the combined prevalence may reach up to 30.4% (4).

This upward trend is partly attributed to increasing maternal age and the implementation of the two-child policy, which have contributed to a growing number of pregnancies affected by both conditions, highlighting the importance of focused perinatal management in this high-risk group (5). Existing studies have identified that factors such as glycemic control levels, mid-pregnancy blood pressure, proteinuria, and a history of preterm birth are closely associated with preterm birth risk (6–9). However, research focusing on the prediction of preterm birth risk in this specific high-risk subgroup of pregnant women with comorbid GDM and HDP remains relatively scarce, with most studies being single-center and small-sample designs (10), lacking external validation, which limits the generalizability and clinical applicability of such models. To date, no prediction models have been specifically developed and externally validated for women with comorbid GDM and HDP.

In addition, although traditional logistic regression models offer good interpretability, they face performance bottlenecks in handling the complex, nonlinear relationships inherent in high-dimensional clinical data (11). In recent years, machine learning algorithms, such as random forest and Extreme Gradient Boosting (XGBoost), have been widely applied in medical prediction studies due to their superior modeling capabilities. Meanwhile, the introduction of interpretability tools such as Shapley Additive Explanations (SHAP) has provided mechanistic explanations for “black-box” models (12), enhancing the clinical interpretability and applicability of these models. By integrating traditional logistic regression and multiple mainstream machine learning algorithms, and systematically evaluating model discrimination, calibration, and clinical utility through receiver operating characteristic (ROC) curves, decision curve analysis (DCA, which evaluates whether using the model provides greater net benefit for clinical decision-making compared with treating all or no patients), and SHAP (which decomposes model predictions to quantify the contribution of each predictor at both the population and individual levels), we aimed to develop an accurate, robust, and interpretable preterm birth risk prediction tool to support early identification and individualized intervention strategies for high-risk pregnancies.

We hypothesized that applying machine learning to GDM–HDP data would yield superior predictive performance for preterm birth compared with traditional models. Therefore, the present study aimed to establish a clinically applicable and interpretable machine learning–based prediction model for preterm birth in women with comorbid GDM and HDP, systematically evaluating its discrimination, calibration, clinical utility, and interpretability.

The development cohort included pregnant women who received antenatal care and delivered at Sichuan Provincial People’s Hospital between January 1, 2020, and December 31, 2024, while the external validation cohort included women who delivered at Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology between January 1, 2022, and December 31, 2023, all of whom met the same inclusion and exclusion criteria. The study was approved by the Ethics Committees of Sichuan Provincial People’s Hospital (No. 2025462) and Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology (No. TJ-IRB20220611), and all data were anonymized and used solely for research purposes.

The inclusion criteria were as follows. (1) Eligible participants were pregnant women aged over 18 years. (2) All participants met the diagnostic criteria for both GDM and HDP according to the guidelines of the Chinese Society of Obstetrics and Gynecology (CSOG) (13, 14). GDM was diagnosed by a 75-g oral glucose tolerance test (OGTT) performed at 24 – 28 gestational weeks if any of the following plasma glucose thresholds were met: fasting ≥5.1 mmol/L, 1-hour ≥10.0 mmol/L, or 2-hour ≥8.5 mmol/L. HDP was diagnosed as systolic blood pressure (SBP) ≥140 mmHg and/or diastolic blood pressure (DBP) ≥90 mmHg after 20 weeks of gestation, confirmed by at least two measurements taken ≥4 hours apart, or a single measurement of SBP ≥160 mmHg and/or DBP ≥110 mmHg, without subtype differentiation. (3) Participants were required to have received continuous and systematic perinatal management in the hospital from early pregnancy (8 – 15 weeks), with no fewer than five prenatal examinations. (4) Only singleton pregnancies with live births were included.

The exclusion criteria were as follows. (1) Pregnancies complicated by severe chronic systemic diseases, such as systemic lupus erythematosus or malignancies, that could affect pregnancy outcomes were excluded. (2) Pregnancies with major fetal malformations were also excluded. (3) Cases with missing key variables that could not be restored through imputation were excluded.

The primary outcome of interest was preterm birth, defined as delivery occurring prior to 37 gestational weeks (15). Outcome data were obtained from the discharge records, labor course records, and ultrasound information in the electronic medical records system and were independently confirmed by two researchers. Continuous variables were tested for normality using the Shapiro–Wilk test. Normally distributed variables were expressed as mean ± standard deviation (SD), and non-normally distributed variables as median with interquartile range (IQR). Categorical variables were summarized as frequencies (percentages). Group comparisons were performed using the t-test or the Mann-Whitney U test for continuous variables and the Chi-square or Fisher’s exact test for categorical variables, as appropriate. To address missing data, the MissForest algorithm—a non-parametric multiple imputation method based on random forests—was applied. All variables had missing values below 10%, which is generally considered acceptable and unlikely to bias the results. This approach iteratively imputes missing values using regression or classification trees trained on observed data, thereby preserving nonlinear relationships among variables. Ten-fold imputation was conducted separately within the development and validation cohorts to prevent information leakage and maintain dataset integrity.

The candidate predictors encompassed several domains: demographic characteristics (Age, body mass index [BMI]) (16); obstetric history (Adverse Pregnancy History and Primiparity) (17, 18); mode of conception (natural conception or in vitro fertilization and embryo transfer [IVF-ET]); pregnancy complications (specifically, the use of antihypertensive medications); and longitudinal measurements of systolic and diastolic blood pressure (SBP and DBP) collected across six gestational intervals: 8 + 0 to 15 + 6, 16 + 0 to 19 + 6, 20 + 0 to 23 + 6, 24 + 0 to 27 + 6, 28 + 0 to 31 + 6, and 32 + 0 to 36 + 6 weeks. Time-specific blood pressure variables were denoted using the format SBP_X_Y or DBP_X_Y, where X_Y indicates the corresponding gestational week range. For example, SBP_32_36 refers to SBP measurements taken between 32 + 0 and 36 + 6 weeks of gestation. Laboratory variables included the mid-pregnancy OGTT (24–28 weeks), with glucose concentrations measured at 0 hours (OGTT-0h), 1 hour (OGTT-1h), and 2 hours (OGTT-2h) after glucose load; liver function markers, including alanine aminotransferase (ALT), aspartate aminotransferase (AST), and lactate dehydrogenase (LDH); as well as Uric Acid, Albumin, and Anemia, totaling more than 30 candidate variables. All variables were collected before the occurrence of outcomes, and outcome data were blinded during data processing to prevent information leakage.

To address the relatively small sample size and the imbalance in outcome distribution, the Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively to the training folds within cross-validation, while validation and test sets remainedunchanged to avoid information leakage. In this dataset, preterm birth cases represented the minority class, whereas non-preterm cases were the majority class; the minority class was oversampled to achieve a 1:1 ratio with the majority class. We optimized SMOTE’s neighborhood parameter (k) using grid search, selecting k = 5. This value was chosen because it achieved the highest overall and negative class F1 scores (the harmonic mean of precision and recall) during a three-fold cross-validation on the development cohort. This strategy enhanced the model’s sensitivity to preterm prediction while minimizing potential bias introduced by synthetic data. A two-step variable selection process was then implemented. First, univariate logistic regression was conducted to screen candidate predictors, and those with a P-value < 0.20 were retained for further modeling, in accordance with Steyerberg’s recommendation in Clinical Prediction Models to preserve variables with potential predictive value (19). Guided by the events-per-variable (EPV) principle, we aimed to maintain a relatively high EPV value to reduce the risk of overfitting given the limited sample size (development cohort: 121 participants, 31 events). To further address the potential impact of a lower EPV in this context, we applied Elastic Net regularization, combined with three-fold cross-validation, to enhance model stability, and conducted external validation and sensitivity analyses to ensure robustness and generalizability. To reduce multicollinearity and avoid overfitting while maintaining a minimum EPV ratio of at least 10 (20), Elastic Net regression—combining the L1 penalty of Least Absolute Shrinkage and Selection Operator (LASSO) and the L2 penalty of Ridge Regression—was applied to identify the most predictive features. Three-fold cross-validation was used to improve model stability. The final model selection was based on cross-validation performance. It is noteworthy that the use of the Naive Bayes (NB) model was pre-specified in our analysis plan, given its advantages in small-sample scenarios and its probabilistic interpretability. It was not chosen post-hoc based on its performance on an external validation set. A total of five key predictors were ultimately retained for final model development.

Model performance was evaluated across multiple dimensions: (1) discrimination was assessed using ROC curves and area under the curve (AUC) values; (2) calibration was evaluated using the Hosmer-Lemeshow test and calibration plots to assess agreement between predicted probabilities and observed outcomes; (3) clinical utility was examined using DCA to estimate net benefit under different threshold probabilities; (4) interpretability was evaluated using SHAP to quantify the direction and contribution of each predictor to individual predictions; (5) generalizability was assessed using an external validation cohort; and (6) reclassification performance was evaluated using integrated discrimination improvement (IDI) and net reclassification improvement (NRI) indices. All statistical analyses were conducted using R (version 4.2.3) and Python (version 3.12), with a two-sided P-value < 0.05 considered statistically significant.

Preterm birth remains a leading contributor to perinatal morbidity and mortality (4), and prediction continues to be particularly challenging in women with comorbid GDM and HDP. Population-based data demonstrate markedly higher risks in GDM–HDP pregnancies compared to either condition alone (21). Prior prediction models have typically addressed only GDM or HDP in isolation, reporting moderate discrimination and limited clinical utility (22). To our knowledge, no existing model specifically targets the combined GDM–HDP population, despite its elevated baseline risk. By focusing on this high-risk group and performing external validation, the present study adds important evidence to this limited domain and proposes a clinically interpretable model that balances predictive performance with translational feasibility. Recent advances in AI and interpretable machine learning further support the feasibility of deploying such models in obstetric practice (23).

Modern ensemble algorithms such as Random Forest and XGBoost have frequently been highlighted for superior discrimination in obstetric prediction (24, 25). However, their complexity and limited interpretability have restricted clinical translation, echoing concerns raised in other obstetric risk prediction tasks, including postpartum hemorrhage (25, 26). In our cohorts, LASSO achieved the highest AUC, and ensembles demonstrated favorable net benefit in decision-curve analyses (27, 28). However, the Naive Bayes (NB) model, while showing a slightly lower AUC (~0.78), achieved superior reclassification (NRI/IDI) and maintained competitive calibration. More importantly, its transparent probabilistic framework provides directly interpretable risk probabilities, which we considered a decisive advantage for practical obstetric risk counseling. This interpretability, combined with the model’s simplicity and routine clinical availability of its predictors, underscores its translational potential in obstetric practice despite the modest AUC. Given the modest sample size and the comparison of multiple algorithms, we acknowledge the potential risk of overfitting and the post hoc nature of model selection. This deliberate trade-off underscores that marginal improvements in AUC may be less clinically meaningful than ensuring interpretability and usability in high-risk care pathways. Accordingly, NB was prioritized for primary reporting, as its simplicity and interpretability enhance translational potential and reduce the risk of overfitting in modest obstetric datasets, aligning with prior calls for clinically explainable prediction models (23).

Beyond model performance, the retained predictors also reflect strong biological plausibility. Elevated hepatic enzymes (ALT, AST) have consistently been associated with preeclampsia severity and adverse maternal outcomes (29–31), and our findings extend this evidence by demonstrating their predictive value in a multimorbidity cohort where hepatocellular injury and systemic inflammation may converge with metabolic stress from GDM (32). Albumin, although underused in prior prediction models, has been linked to endothelial dysfunction, maternal malnutrition, and fluid imbalance (32), and its inclusion here emphasizes the interplay between hepatic reserve and vascular integrity in dual-risk pregnancies. Similarly, LDH has long been reported as a marker of cellular injury and oxidative stress in severe HDP (33). While earlier studies positioned LDH primarily as a late indicator of disease severity, our results suggest a broader role in the GDM–HDP setting, integrating systemic hypoxic stress with metabolic dysregulation to capture maternal–placental strain more comprehensively. Compared with organ-specific enzymes, LDH reflects systemic cascades, consistent with recent arguments on systemic predictors of pregnancy complications (35).

Hemodynamic adaptation in late gestation further reinforces this multi-domain perspective. Blood pressure has long been recognized as central to pregnancy outcomes (35), yet most prior prediction models have focused on early-pregnancy measures for anticipatory stratification (36). Our identification of systolic blood pressure at 32 – 36 weeks (SBP_32_36) as an independent predictor highlights the prognostic significance of late-gestational dynamics. Elevated SBP in this window likely reflects cumulative vascular burden and declining compensatory capacity, suggesting that temporal patterns of blood pressure provide additional prognostic information. By incorporating such longitudinal measures, our approach moves prediction closer to real-time surveillance, consistent with precision obstetric care initiatives (37).

Taken together, these findings both corroborate and extend existing literature. They confirm the roles of hepatic enzymes (29–31), LDH (34), and blood pressure trajectories (35), but in a broader context that integrates multimorbidity rather than single conditions. By situating individual predictors within the dual-risk framework of GDM–HDP, our study highlights not only their continued relevance but also new dimensions, such as the overlooked predictive value of late-gestational SBP. Divergences from prior reports likely reflect differences in study populations, sample sizes, and timing of data collection, but they also underscore the distinctive pathophysiology of multimorbidity. Thus, our study provides a more integrated and clinically relevant framework for preterm birth prediction. Importantly, the interpretability of the model strengthens its applicability in clinical practice. SHAP-based visualizations at both population and individual levels allowed us to bridge predictive accuracy with explainability (38), a recognized barrier to adoption of Artificial Intelligence in obstetrics. SHAP summary plots consistently ranked LDH, albumin, and SBP_32_36 among the most influential predictors, reinforcing their biological plausibility and echoing recent applications of SHAP in preeclampsia and postpartum hemorrhage models. Notably, SHAP plots also suggested a tendency to underestimate risk among the lowest-risk strata, which may reflect data imbalance and calibration limitations. In parallel, patient-level SHAP force plots decomposed individual risk profiles into positive and negative contributions, offering a pathway to targeted monitoring and intervention. In addition, we leveraged the intrinsic transparency of the Naive Bayes model itself, whose conditional probabilities directly reflect the contribution of each predictor to the overall risk. This NB-specific interpretability complements the SHAP explanations, providing clinicians with a more intuitive understanding of risk attribution. Although SHAP may introduce approximation errors and relies on assumptions of feature independence, these limitations are mitigated within the simple probabilistic structure of NB.

Clinical relevance further underscores the potential utility of our model. Because all five predictors are routinely measured in antenatal care, the model can be seamlessly applied in real-world workflows without additional testing burden. In particular, SBP measured at 32 – 36 weeks provides a practical window to inform delivery planning and closer surveillance. While the model may add limited value in cases of overtly severe GDM or HDP, it offers important guidance in borderline or ambiguous presentations, helping clinicians to stratify risk more objectively. Moreover, by integrating multiple predictors, the model can reveal hidden risk profiles that may not be apparent when considering single variables in isolation. This potential to enhance workflow efficiency and provide early warnings highlights its clinical value. However, the model should be viewed as a decision-support tool that may aid in risk stratification and clinical management, rather than one that directly improves outcomes.

To evaluate real-world robustness, we also assessed the impact of class imbalance. The rarity of adverse outcomes often limits model performance, and prior work has noted the instability of ensemble methods under imbalance (39). In our analyses, the NB model maintained stable calibration and discrimination following SMOTE while improving recall. Sensitivity analyses further indicated that SMOTE improved recall and balanced accuracy without substantially altering the AUC or calibration, suggesting that the oversampling procedure effectively mitigated class imbalance without introducing significant bias. Comparable trends were observed without SMOTE, albeit with slightly reduced discrimination, supporting the robustness of the findings. This indicates that NB, combined with oversampling, may be particularly suitable for obstetric applications where event rates are low and interpretability is essential.

Finally, strengths and limitations warrant consideration. This is, to our knowledge, the first study to develop and externally validate a preterm birth prediction model specifically for women with comorbid GDM and HDP. Strengths include the use of dual-center data to enhance generalizability; the integration of five clinically accessible, biologically coherent predictors spanning complementary physiological domains; variable selection through Elastic Net to address multicollinearity, followed by a robust yet simple NB classifier; and a comprehensive multi-metric evaluation covering discrimination, calibration, clinical utility, reclassification, and interpretability. Limitations include its retrospective design, the relatively small sample size (particularly in the external validation cohort), potential selection bias, the absence of intervention-based validation, and the possibility that the findings may not be directly generalizable to populations outside China. In addition, we recognize the possibility of overfitting due to modest sample size and the post hoc nature of model choice. Moreover, some clinically relevant domains such as psychosocial, nutritional, or imaging biomarkers were not included. Medication variables were also coarse, with antihypertensive use captured only as a binary yes/no indicator and glucose-lowering therapies not systematically recorded, which may limit interpretability. Future research should pursue prospective and multicenter validation, explore integration into clinical workflows via electronic health records or mobile/desktop applications, and evaluate whether model-guided interventions can improve maternal and neonatal outcomes.

In this study, we developed and externally validated multiple predictive models to assess the risk of preterm birth in pregnancies complicated by both GDM and HDP. Among the evaluated algorithms, the Naive Bayes classifier demonstrated the most favorable balance across discrimination, reclassification, interpretability, and robustness, and was ultimately selected as the optimal model for clinical application. Through Elastic Net regression, five physiologically meaningful predictors—ALT, AST, albumin, LDH, and systolic blood pressure at 32 – 36 weeks—were identified and incorporated into model development. These variables capture distinct domains relevant to preterm labor pathophysiology, including hepatic dysfunction, systemic inflammation, vascular insufficiency, and hemodynamic instability. To enhance transparency and clinical utility, SHAP-based interpretation techniques were applied at both the global and individual levels. Summary plots highlighted the dominant predictors at the population level, while force plots provided case-specific insights into individualized risk contributions. Additionally, SMOTE-based sensitivity analysis confirmed the Naive Bayes model’s robustness under class imbalance, further supporting its generalizability and deployment potential.

The proposed Naive Bayes model may assist clinicians in early identification and personalized risk management of high-risk pregnancies affected by GDM and HDP, and represents a step toward the implementation of transparent, evidence-based decision support in obstetric practice. Future studies should aim to validate this model in larger, multicenter cohorts and explore its integration into real-time clinical decision support systems.