Comparison of Interpretable Machine Learning Models Using Systemic Inflammation Index to Predict Preterm Birth in Gestational Diabetes Mellitus

Introduction

Gestational diabetes mellitus (GDM) is a common metabolic disorder characterized by carbohydrate intolerance and represents one of the fastest-growing pregnancy complications. It is defined as the first occurrence or detection of abnormal glucose metabolism during pregnancy.^1,2 Beyond its metabolic implications, this condition significantly elevates the risk of adverse perinatal outcomes, such as preterm birth (<37 weeks’ gestation).^3,4 Although current management combining lifestyle modifications and pharmacological therapy can improve glycemic control and reduce preterm birth risk, a clinically significant proportion of GDM patients still experience preterm delivery.^5–7 Preterm birth imposes a dual burden, constituting both a major financial strain on families and the leading cause of under-5 mortality worldwide.⁸ Early prediction and prevention of preterm delivery in GDM patients could significantly improve perinatal outcomes while alleviating healthcare system pressures. This underscores the critical need for reliable preterm birth prediction methods in this high-risk population.

GDM pathogenesis demonstrates significant associations with both established clinical predictors (including maternal age ≥35 years, smoking status, and insulin resistance)^9,10 and inflammatory alterations.^11,12 Although current prediction models based on clinical parameters achieve only modest sensitivity (51.5%) for preterm birth,¹³ the incorporation of routinely measured inflammatory markers (such as lymphocyte and neutrophil counts) may enhance predictive accuracy. In GDM, circulating inflammatory cells (such as monocytes and neutrophils) and proinflammatory cytokines (eg, IL-1β, IL-6, TNF-α) are elevated. As pregnancy progresses, increasing placental and adipose tissue further boosts cytokine secretion, while elevated glucose levels activate inflammatory pathways, leading to the release of additional cytokines and chemokines (eg, CXCL1, CXCL5, CXCL8, CCL2).¹⁴ Parturition involves uterine contractions, cervical dilation, and membrane rupture, all driven by inflammation.^15–20 GDM may prematurely trigger this inflammatory process, increasing the risk of preterm birth. Recent studies highlight the prognostic value of hematologic inflammatory indices - including neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), lymphocyte-to-monocyte ratio (LMR), and systemic immune-inflammation index (SII) - in assessing systemic inflammation and disease severity.²¹ Higher levels of NLR, PLR, LMR, and SII are associated with GDM.²² Derived from routine complete blood count parameters (neutrophils, lymphocytes, and monocytes), these indices demonstrate significant clinical advantages: they are cost-efficient, easily obtainable, and provide comprehensive profiling. Importantly, their potential association with preterm birth pathogenesis suggests these biomarkers may improve risk prediction in GDM patients. Nevertheless, the complex, nonlinear relationships between multidimensional clinical data and patient outcomes present significant challenges for conventional linear models (eg, logistic regression [Logit]), limiting their predictive accuracy. In this context, machine learning (ML) methods offer distinct advantages due to their capacity to identify sophisticated patterns in high-dimensional biomedical data.²³ Although ML has demonstrated success in various medical domains, including oncology²⁴ and cardiology,²⁵ its application for preterm birth prediction in GDM using systemic inflammation biomarkers and traditional clinical predictors remains underexplored.

In this context, this study aims to develop and validate an interpretable ML model that integrates systemic inflammation indices with traditional clinical predictors to identify novel and practical obstetric biomarkers for preterm birth risk assessment in GDM patients. By enabling early risk stratification at the time of GDM diagnosis, this approach will facilitate timely clinical interventions to prevent preterm delivery in this high-risk population.

Methods

Ethics Statement

This study received approval from the medical ethics committee of Shanghai University of Medicine & Health Sciences affiliated Zhoupu Hospital (Approval No. 2024-C-160-E01) and complied with the Declaration of Helsinki. Given the retrospective, observational design, the ethics committee granted a waiver of individual informed consent. Patient confidentiality was maintained through comprehensive deidentification procedures, with systematic removal of all personal identifiers from electronic health records prior to analysis in accordance with institutional privacy standards.

Subjects and Study Design

This retrospective cohort study was conducted from August 2019 to August 2024, during which 568 patients with GDM were consecutively screened from Shanghai University of Medicine & Health Sciences affiliated Zhoupu Hospital. Participants met the following criteria: (1) age ≥18 years; (2) GDM diagnosis confirmed by 75g oral glucose tolerance test (OGTT) at 24–28 weeks’ gestation (fasting glucose ≥5.1 mmol/L, 1-hour ≥10.0 mmol/L, or 2-hour ≥8.5 mmol/L); and (3) singleton pregnancy. Exclusion criteria included pre-existing diabetes, multifetal gestation, prior preterm birth, incomplete medical records, and significant comorbidities (renal, hepatic, or cardiovascular diseases). Following strict inclusion/exclusion criteria, 389 GDM patients were enrolled and divided into two cohorts: (1) a training cohort (n=272; August 2019 to March 2023), which included subgroups of term birth and preterm birth; and (2) a temporal external validation cohort (n=117; April 2023 to August 2024) (Figure 1A–C).

Figure 1 Flowchart of patient selection and cohort distribution for developing and validating predictive models to assess preterm birth risk in patients with GDM. Abbreviations: GDM, gestational diabetes mellitus; ML, machine learning; SHAP, Shapley Additive Explanations.

Data Collection

This study collected comprehensive maternal data including demographic characteristics (age, pre-pregnancy BMI), medical history (diabetes family history, smoking/alcohol use, hypertension, parity, uterine curettage, IVF-ET), OGTT results, and hematological indices at GDM diagnosis (complete blood counts and derived inflammatory ratios: NLR, PLR, LMR, SII).

Feature Preprocessing and Selection in the Training Cohort

Before model development, data preprocessing was performed to ensure fairness. Continuous variables were standardized to Z-scores (mean = 0, standard deviation = 1), and categorical variables were binarized to “0” or “1,” ensuring fair comparison across scales.

To simplify the model, we first identified statistically significant variables distinguishing between term and preterm birth groups using univariate logistic regression, followed by appropriate statistical tests (Student’s t-test, Mann–Whitney U-test, or chi-square test based on data distribution) to validate the results of the univariate logistic regression. Variables with P < 0.05 in the univariate analysis were then included in a multivariable logistic regression to identify independent risk factors and construct the predictive model.

Development and Internal Validation of Prediction Models

In our study, a marked class imbalance was observed, with far fewer preterm births than term births. Such disparities are common in medical datasets, where the prevalence of non-cases typically outweighs cases, often impairing predictive accuracy.²⁶ To address this issue, we applied the Synthetic Minority Oversampling Technique (SMOTE), which generates synthetic samples based on k-nearest neighbors to balance minority and majority classes.^27,28 This method has been shown to enhance disease prediction and reduce model overfitting. Importantly, SMOTE was implemented only in the training cohort, while the validation cohort remained untouched to preserve its natural outcome distribution.

We developed four ML models (Logit, random forest [RF], support vector machines [SVM], and extreme gradient boosting [XGBoost]) to predict preterm birth risk in GDM patients using: (1) traditional clinical parameters, (2) systemic inflammatory markers, and (3) their combination. A triple five-fold cross-validation strategy prevented model overfitting. A nested cross-validation framework was employed during model development, using stratified 5-fold cross-validation to preserve outcome distribution and evaluate generalization performance. The outer loop assessed performance, while the inner loop, optimized through grid search, focused on hyperparameter tuning to prevent data leakage. This process was repeated across 1000 bootstrap iterations. Implemented with the StratifiedKFold function in scikit-learn, it separates model selection from evaluation, enhancing generalization and reducing overfitting risk.

Hyperparameter optimization for the ML models was performed using randomized grid search with 5-fold cross-validation to avoid information leakage. The LR grid included regularization strength and penalty type; RF focused on tree depth, estimators, and minimum samples per leaf; XGBoost tuned learning rate, max depth, and estimators; and SVM adjusted kernel, regularization, and kernel coefficient. Each model underwent 100 iterations, with area under the receiver operating characteristic curve (AUC-ROC) as the optimization metric. After identifying optimal parameters, final models were trained on the training cohort to maximize training data use.

Temporal External Validation and Interpretability of ML Models

The model’s performance was evaluated using discrimination metrics (AUC-ROC, area under the precision-recall curve [AUC-PRC], sensitivity, specificity, positive predictive value [PPV], negative predictive value [NPV], F1 score), calibration measures (brier score, calibration curves), and clinical utility assessment via decision curve analysis (DCA). The Brier score quantifies prediction accuracy (lower values indicating better calibration), while DCA estimates net clinical benefit.

Based on cooperative game theory,²⁹ SHAP (SHapley Additive exPlanations) analysis quantifies variable importance through Shapley values, representing each feature’s predictive contribution. This method provides both: (i) quantitative assessment of directional feature effects (protective/risk factors), and (ii) visual interpretation via summary plots (population-level importance) and force plots (individual-case predictions).^30–32

Statistical Analysis

We conducted additional sensitivity analyses to strengthen the reliability of our findings. First, we excluded participants with a history of hypertension, smoking, or alcohol use at baseline to minimize potential confounding. Second, to address missing data, we used multiple imputation by chained equations (MICE), which preserved statistical power and minimized bias associated with missing values.

Data were analyzed using IBM SPSS (v26.0), R (v4.2.3), and Python (v3.10.0). Continuous variables were assessed for normality via the Shapiro–Wilk test and summarized as mean ± SD (normal distribution) or median (interquartile range [IQR]) (non-normal distribution), with between-group comparisons conducted using Student’s t-test or Mann–Whitney U-test, respectively. Categorical variables were presented as n (%) and compared using χ²-tests. Statistical significance was set at p < 0.05 (two-tailed).

Results

Patient Characteristics

Figure 1A–C outlines the participant selection process, with 389 GDM patients (of 568 screened) meeting eligibility criteria. The preterm birth prevalence was similar between cohorts: 13.2% (36/272) in the training set versus 14.5% (17/117) in the temporal external validation set (χ²=0.077, p=0.782). Table 1 demonstrates comparable baseline characteristics between cohorts, including traditional clinical parameters and systemic inflammation indices (all p>0.05).

Table 1 Baseline Characteristics of Patients in the Training and Validation Cohorts

Feature Selection in the Training Cohort

Table 2 presents the logistic regression analyses identifying preterm birth risk factors in the training cohort. Univariate analysis revealed significant associations (P<0.05) for eight variables: maternal age, uterine curettage, and six inflammatory markers (neutrophil count, monocyte count, NLR, PLR, LMR, SII), with consistent non-parametric test results (Table S1). Multivariate analysis identified seven independent predictors: maternal age, uterine curettage, and five inflammatory indices (monocyte count, NLR, PLR, LMR, and SII) (all P<0.05; Table 2). Multicollinearity was assessed using the variance inflation factor (VIF), with all independent predictors showing VIF values below 2, indicating no multicollinearity. Therefore, LASSO regression is likely unnecessary. These seven key variables were subsequently incorporated into the final ML model. These inflammatory indices can be derived from routine blood tests, making them practical, rapid, and cost-effective tools. This is particularly advantageous in healthcare settings with limited resources, where access to advanced imaging or molecular diagnostics may be restricted.

Table 2 Univariate and Multivariate Logistic Analyses to Determine the Independent Predictors Associated with Preterm Birth in the Training Cohort

Comparing Models for Predicting Preterm Birth Risk

We evaluated four ML algorithms (Logit, SVM, RF, and XGBoost) for preterm birth prediction in GDM patients using three predictor sets: traditional clinical parameters, systemic inflammation indices, and their combined features. The combined-feature models demonstrated superior performance (AUC-ROC 0.847–0.932; AUC-PRC 0.582–0.754) versus clinical-only (AUC-ROC 0.761–0.820; AUC-PRC 0.394–0.486) or inflammation-only (AUC-ROC 0.786–0.872; AUC-PRC 0.475–0.708) approaches (DeLong’s test, p<0.05), as shown in Table 3 (performance metrics) and Figures 2–4 (ROC, PRC, calibration, and decision curves).

Table 3 Performance of ML Classifiers in Predicting Preterm Birth Risk in GDM Using Traditional Clinical Data, Systemic Inflammation Markers, and Combined Datasets

Figure 2 Performance comparison of ML classifiers (Logit, SVM, RF, XGBoost) on clinical data: (A) ROC curves (AUC-ROC: 0.761–0.820), (B) precision-recall curve (AUC-PRC: 0.394–0.486), (C) calibration plots, and (D) DCA. Abbreviations: ML, machine learning; AUC-ROC, area under the receiver operating characteristic curve; AUC-PRC, area under the precision-recall curve; DCA, decision curve analysis; Logit, logistic regression; SVM, support vector machine; RF, random forest; XGBoost, extreme gradient boosting.

Figure 3 Performance comparison of ML classifiers (Logit, SVM, RF, XGBoost) on systemic inflammation index: (A) ROC curves (AUC-ROC: 0.786–0.872), (B) precision-recall curve (AUC-PRC: 0.475–0.708), (C) calibration plots, and (D) DCA. Abbreviations: ML, machine learning; AUC-ROC, area under the receiver operating characteristic curve; AUC-PRC, area under the precision-recall curve; DCA, decision curve analysis; Logit, logistic regression; SVM, support vector machine; RF, random forest; XGBoost, extreme gradient boosting.

Figure 4 ML classifiers (Logit, SVM, RF, XGBoost) on combined traditional clinical and systemic inflammation data: (A) ROC curves (AUC-ROC: 0.847–0.932), (B) precision-recall curve (AUC-PRC: 0.582–0.754), (C) calibration, (D) DCA. Abbreviations: ML, machine learning; AUC-ROC, area under the receiver operating characteristic curve; AUC-PRC, area under the precision-recall curve; DCA, decision curve analysis; Logit, logistic regression; SVM, support vector machine; RF, random forest; XGBoost, extreme gradient boosting.

XGBoost demonstrated superior predictive performance among combined clinical-inflammatory models, achieving the highest discriminative accuracy (AUC-ROC = 0.932; AUC-PRC=0.754) with excellent calibration, particularly in the clinically relevant risk range below 30%. DCA confirmed robust clinical utility across all models, with XGBoost consistently outperforming in sensitivity, specificity, PPV, NPV, F1-score, and Brier score. These findings support XGBoost as the preferred model for preterm birth risk prediction in GDM patients. Furthermore, our findings align with a previous study indicating that ML algorithms often outperform traditional linear approaches in terms of accuracy.³³

Temporal External Validation of ML Model Performance

The XGBoost model’s predictive performance was externally validated using ROC analysis (AUC-ROC=0.893, AUC-PRC=0.667; Figures 5A and B), calibration curves (Figure 5C), and DCA (Figure 5D). While showing marginally reduced accuracy compared to the training set, the model demonstrated: (1) maintained discriminative capacity, (2) excellent prediction-observation concordance, and (3) clinically meaningful net benefits. These results confirm the model’s robustness for preterm birth risk stratification in GDM patients.

Figure 5 Validation of the optimal ML model in a temporal external cohort: (A) ROC (AUC-ROC=0.893), (B) precision-recall curve (AUC-PRC=0.667), (C) calibration, (D) DCA. Abbreviations: ML, machine learning; AUC-ROC, area under the receiver operating characteristic curve; AUC-PRC, area under the precision-recall curve; DCA, decision curve analysis.

Interpretation of the Model

To evaluate feature importance in the XGBoost model, we applied SHAP analysis, which quantifies each predictor’s contribution using absolute mean SHAP values. This revealed the top five influential factors: two clinical parameters (maternal age and uterine curettage history) and three systemic inflammation markers (SII, PLR, and NLR) (Figure 6). The SHAP summary plot (Figure 6A) displays individual feature impacts across patients, with point colors indicating feature values (yellow = high, blue = low). The horizontal axis represents SHAP values, with feature importance visually emphasized by larger data point clusters - the wider the spread, the stronger the predictive influence on preterm birth risk in GDM patients. Complementing this, the importance bar chart (Figure 6B) ranks variables by their overall predictive power. Key features, in descending order of significance, were: SII, PLR, maternal age, NLR, uterine curettage history, monocyte count, and LMR. Notably, inflammatory indices dominated the top predictors, underscoring their clinical relevance.

Figure 6 SHAP analysis of XGBoost for preterm birth prediction: (A) summary plot, (B) feature importance. Abbreviations: SHAP, Shapley Additive Explanations; XGBoost, extreme gradient boosting; SII, systemic immune-inflammation index; PLR, platelet-to-lymphocyte ratio; NLR, neutrophil-to-lymphocyte ratio; LMR, lymphocyte-to-monocyte ratio.

The SHAP force plot (Figure 7) visualizes feature contributions to individual patient predictions, where yellow and red regions respectively indicate risk-increasing and protective factors for preterm birth in GDM patients, with region width reflecting effect magnitude. The output value f(x) aggregates all feature contributions (SHAP values) for a given patient, while the base value represents the population average prediction. The SHAP force plot demonstrates XGBoost’s predictive accuracy through two representative cases: the upper panel (Figure 7A) correctly predicts preterm birth based on elevated SII and PLR values, advanced maternal age, and other contributing factors, while the lower panel (Figure 7B) accurately identifies a term birth case characterized by lower SII, younger maternal age, and additional protective indicators. This method enables precise differentiation between the risks of preterm birth and term birth, offering personalized risk assessments for GDM patients and facilitating early clinical intervention.

Figure 7 SHAP force plots for individual predictions: (A) preterm birth case (B) term birth case. Abbreviations: SHAP, Shapley Additive Explanations; SII, systemic immune-inflammation index; PLR, platelet-to-lymphocyte ratio; NLR, neutrophil-to-lymphocyte ratio.

Sensitivity Analyses

Sensitivity analyses confirmed the robustness of our findings. After excluding participants with a history of hypertension, smoking, or alcohol use, the XGBoost model demonstrated performance consistent with the primary analysis in the combined clinical-inflammatory models (AUC-ROC=0.891, AUC-PRC=0.670, Figures 8A–D). Furthermore, the use of MICE for handling missing data, compared with complete-case analysis, did not materially affect model performance (AUC-ROC=0.893, AUC-PRC=0.620, Figures 9A–D).

Figure 8 Sensitivity analyses were performed in participants without a history of hypertension, smoking, or alcohol use at baseline (N=330). The XGBoost model demonstrated performance consistent with the primary combined clinical-inflammatory analysis, as assessed by (A) ROC curve (AUC-ROC=0.891), (B) precision-recall curve (AUC-PRC=0.670), (C) calibration plots, and (D) DCA. Abbreviations: AUC-ROC, area under the receiver operating characteristic curve; AUC-PRC, area under the precision-recall curve; DCA, decision curve analysis; XGBoost, extreme gradient boosting.

Figure 9 Sensitivity analyses were performed after imputation of missing data (N=405). Compared with the complete-case analysis, imputation did not materially affect the performance of the XGBoost model, as shown by (A) ROC curve (AUC-ROC=0.893), (B) precision-recall curve (AUC-PRC=0.620), (C) calibration plots, and (D) DCA. Abbreviations: AUC-ROC, area under the receiver operating characteristic curve; AUC-PRC, area under the precision-recall curve; DCA, decision curve analysis; XGBoost, extreme gradient boosting.

Discussion

GDM, a prevalent metabolic disorder of pregnancy, significantly increases preterm birth risk and associated adverse outcomes despite current management strategies.^7,34,35 Developing reliable prediction methods for preterm birth in GDM patients is crucial to improve perinatal outcomes and reduce healthcare burdens. GDM pathogenesis is significantly associated with both established clinical predictors (such as maternal age ≥35 years and smoking status)^9,36 and inflammatory biomarkers,^11,12,22,37 and may enhance the modest predictive accuracy of current preterm birth models, which currently neglect some inflammatory biomarkers like PLR, LMR, SII.¹³ While conventional linear models (eg, Logit) struggle with the complex nonlinear relationships between multidimensional clinical data and patient outcomes, ML methods remain underexplored for preterm birth prediction in GDM despite their superior ability to decipher sophisticated patterns in combined clinical and systemic inflammation biomarker data. Our study addressed this critical gap by developing and comparing four ML algorithms that synergistically combined conventional clinical factors with systemic inflammatory markers. The XGBoost model demonstrated superior performance in predicting preterm birth among GDM patients by effectively integrating both data types. Through SHAP value analysis, we enhanced model interpretability and gained novel insights into the contribution of inflammatory pathways to preterm birth risk. Sensitivity analyses further confirmed the reliability of these results. These findings advance the field by establishing an interpretable ML framework that combines traditional clinical and systemic inflammatory markers for early risk stratification, enabling more targeted interventions in GDM management.

Our study employed ML algorithms to overcome the limitations of conventional linear models in analyzing complex non-linear relationships.³⁸ We evaluated four ML models comparing three approaches: clinical indicators alone, systemic inflammation indices alone, and their combined integration. The integrated models achieved superior predictive performance for preterm birth, as their multidimensional analysis captures synergistic interactions between clinical and inflammatory factors that singular approaches miss. This comprehensive integration strategy enhanced prediction accuracy.

XGBoost emerged as the optimal model in our evaluation, demonstrating consistent high accuracy in temporal external validation through its integration of traditional clinical indicators with systemic inflammation indices. Li et al developed a nomogram for predicting preterm birth in pregnant women with GDM using clinical risk factors, but its performance (AUC = 0.722) was notably lower than that of our model (AUC = 0.932).³ While prior research has primarily employed linear models using traditional clinical and inflammatory markers, our study advances the field by applying ML to analyze a broader spectrum of inflammatory biomarkers, such as PLR, LMR, and SII, significantly enhancing prediction accuracy for preterm birth in GDM patients (AUC=0.932 vs 0.885).³⁹ To improve interpretability of the complex ML model, we used SHAP analysis. The SHAP feature importance map visually shows each feature’s impact on model outputs through SHAP values, representing their influence range and direction.⁴⁰ Each plotted point corresponds to a sample, with colored bars indicating feature values and their distribution. Left-positioned bars denote negative impacts, while right-positioned bars show positive effects.^41,42 This approach helps identify key features for model optimization and selection. The SHAP analysis identified five key preterm birth predictors: three systemic inflammation indices and two traditional clinical indicators. Our analysis identified elevated NLR, PLR, and SII levels as significant predictors of preterm birth in GDM patients. These composite inflammatory indices demonstrate greater clinical reliability than isolated hematological parameters (including neutrophil, lymphocyte, and monocyte counts) due to their stability against physiological fluctuations, pathological interferences, and technical measurement variations.^43,44 The SII demonstrates particular clinical value by quantitatively integrating the synergistic interactions among platelets, neutrophils, and lymphocytes.⁴⁵ This composite index may provide a more comprehensive assessment of inflammatory-immune crosstalk compared to isolated ratios like NLR or PLR. In addition to systemic inflammation indices, our study identified maternal age and uterine curettage as additional preterm birth risk factors. Global trends show increasing maternal age, a recognized independent risk factor for GDM. Meta-analyses indicate 2-fold and 4-fold higher GDM risks for women aged 35–39 and >40 years, respectively, compared to those <35 years.^46–50 Additionally, women aged 35–39 face 1.4% preterm birth risk.⁵¹ Building on prior research, our study confirmed maternal age as an independent preterm birth risk factor in GDM patients, enabling better pregnancy outcome assessment. Uterine curettage, particularly for miscarriage or pregnancy termination, is an independent preterm birth risk factor due to potential cervical trauma.⁵² This underscores the need to document termination procedures and evaluate their preterm birth risks. The SHAP-interpreted XGBoost model effectively identified and quantified key predictive factors for preterm birth in GDM patients, providing clinically valuable risk stratification. This ML approach facilitates timely preventive interventions, such as intensive monitoring, corticosteroid administration, or transfer to a higher-level care facility. While these measures improve outcomes, they also carry risks, including unnecessary treatments or hospitalizations for patients who do not deliver preterm.

While demonstrating the predictive potential of ML models combining traditional clinical indicators and systemic inflammation indices for preterm birth in GDM, this study has two main limitations. First, the single-center design with 389 participants may limit generalizability due to potential regional biases, the relatively small sample size, spectrum bias, and limited transportability. Second, the retrospective nature and exclusion of cases with incomplete records could introduce selection bias. These constraints notwithstanding, our findings underscore the clinical value of this integrated approach. Future multi-center prospective studies with larger cohorts are warranted to validate and optimize the model’s performance, ideally including external or geographic validation from additional centers.

In summary, the XGBoost model demonstrates superior predictive performance for preterm birth in GDM patients by synergistically combining traditional clinical parameters with systemic inflammation indices. This integrated ML approach enables precise differentiation between the risks of preterm birth and term birth, offering personalized risk assessments for GDM patients. It supports clinical decision-making, facilitates early intervention, and has the potential to enhance perinatal outcomes.