Approved by the Institutional Ethics Committee of Wuxi People’s Hospital Affiliated to Nanjing Medical University (Approval No. 2021-1624011), this prospective cohort enrolled 50 GDM patients between February 2021 and January 2023. Participants met diagnostic criteria per the 2022 Guidelines for Diagnosis and Management of Hyperglycemia in Pregnancy [16], with gestational ages of 28⁺⁰ to 38⁺⁶ weeks scheduled for elective delivery. Inclusion criteria comprised: (1) completion of AI-based FLM assessment and uterine artery Doppler within 48 hours before delivery; (2) written informed consent. Exclusion criteria included: (1) suboptimal ultrasound image quality; (2) concurrent preeclampsia, fetal anomalies, or major complications; (3) requirement for emergency cesarean delivery; (4) incomplete glucose monitoring records. After applying these criteria, all 50 initially enrolled patients were included in the final analytical cohort, which consisted of 22 NRDS cases and 28 controls. The cohort (age range: 24–38 years; mean gestational age: 35.2 $\pm$ 2.1 weeks) showed 44.0% proportion of NRDS cases (22/50). Glycemic control classification adhered to the 2022 American Diabetes Association guidelines [17], defining optimal control as fasting glucose 5.3 mmol/L combined with 1-hour postprandial glucose 7.8 mmol/L, while suboptimal control encompassed either fasting glucose $\ge$5.3 mmol/L, 1-hour postprandial glucose $\ge$7.8 mmol/L, or requirement for insulin therapy. Among all 50 participants, 18 (36.0%) had suboptimal glycemic control and 32 (64.0%) had optimal control. The suboptimal glycemic group had 13 NRDS cases (72.2%, 13/18) vs. 9 cases in the optimal group (28.1%, 9/32), accompanied by elevated PI values (median 1.61 vs. 1.21, p 0.001), confirming a 2.57-fold increased NRDS risk and hyperglycemia as the primary driver of the metabolic-vascular-pulmonary cascade. The study adhered to the Declaration of Helsinki.

The detailed baseline clinical and biochemical characteristics of the study cohort are provided in Supplementary Material 1.

Transvaginal measurements (PLT-705BT probe, 5–9 MHz) acquired uterine artery spectra at the cervical internal os level. Pulse Doppler settings maintained a 2-mm sample volume and $\le$30° insonation angle. Mean bilateral PI values were calculated from three consecutive waveforms per the International Society of Ultrasound in Obstetrics and Gynecology (ISUOG) standards. All PI values underwent blinded dual-analyst verification with daily device calibration.

The sample size was calculated a priori using GPower 3.1 software (version 3.1.9.7; Heinrich Heine University Düsseldorf, Düsseldorf, Germany) based on expected group differences in uterine artery PI. Assuming a Cohen’s d effect size of 0.8 (derived from prior GDM studies reporting median PI differences $\ge$0.35 SD units), with $\alpha$ = 0.050 (two-tailed) and $\beta$ = 0.200 (80.0% power), the Mann-Whitney U test required 26 participants (13 per group). However, a larger sample size is recommended for the stable training of machine learning models, particularly given the high-dimensional nature of our feature set (342 initial features). Therefore, we prospectively enrolled 50 patients. Continuous variables violating normality (Shapiro-Wilk test p 0.050) were expressed as median (IQR) and analyzed via Mann-Whitney U test. Effect sizes (r = Z/$\surd$N) were interpreted per Cohen’s criteria: r 0.30 (weak), 0.30 $\le$ r 0.50 (moderate), r $\ge$0.5 (strong). Categorical variables underwent Fisher’s exact test. The diagnosis of NRDS was retrospectively established by a blinded neonatal specialist, strictly adhering to standard clinical, radiological, and laboratory criteria. NRDS was defined by the presence of clinical signs of respiratory distress, including tachypnea (respiratory rate $>$70/min), nasal flaring, intercostal or subcostal retractions, and grunting; concurrent with an increased oxygen requirement (FiO₂$>$30% to maintain oxygen saturation $>$90% and/or need for non-invasive or invasive mechanical ventilation); and supported by radiological evidence on chest X-ray (e.g., reticulogranular patterns, air bronchograms, reduced lung volumes) or lung ultrasound findings. Additionally, hypoxemia (PaO₂60 mmHg) and/or hypercarbia (PaCO₂$>$50 mmHg) with normal lactic acid levels (excluding metabolic acidosis) were required. Any neonate receiving exogenous surfactant therapy was classified into the NRDS group. Notably, NRDS is a postnatal diagnosis and was not based on Apgar scores or umbilical cord blood gas analysis, which reflect intrauterine condition. κ assessed AI-clinical diagnosis concordance (Landis-Koch criteria). Decision curve analysis (DCA) computed net benefit (NB) across threshold probabilities (P_t = 10.0%–50.0%) using: NB = TP/N - (FP/N) $\times$ [P_t/(1 - P_t)], where N = 50 (full cohort size), and true positives (TP) represented AI-high-risk NRDS cases. McNemar tests compared diagnostic accuracies, with continuity correction for expected frequencies 5. Analyses used SPSS 26.0 (IBM Corp., Armonk, NY, USA), R 4.3.1 (R Foundation for Statistical Computing, Vienna, Austria) (mda package), and TensorFlow 2.15 LTS. Statistical significance threshold was $\alpha$ = 0.050 (two-tailed).

Diagnostic performance metrics including sensitivity, specificity, and κ were calculated using contingency tables derived from 10-fold cross-validation. It is important to note the distinct purposes of the statistical corrections applied. The Benjamini-Hochberg procedure was employed during the initial, high-dimensional feature screening phase (as described in Section 2.2.2) to control the false discovery rate when evaluating 342 individual GLCM features. In contrast, the subsequent hypothesis tests (e.g., Mann-Whitney U tests for group comparisons) and model performance evaluations were conducted on a consolidated set of variables (the final model output or established clinical parameters). These latter analyses do not involve the same multiplicity issue as the initial feature screening, and therefore report uncorrected p-values to avoid excessive reduction in statistical power for these confirmatory tests. For the key primary outcome (comparison of uterine artery PI between NRDS and control groups), a post-hoc Bonferroni correction was also applied to confirm the robustness of the finding, given its central role in our hypothesis. Analyses used SPSS 26.0, R 4.3.1, and TensorFlow 2.15 LTS with CUDA 11.8 acceleration.

Following a comprehensive retrospective review of neonatal medical records applying the strict clinical, radiological, and laboratory criteria for NRDS (as defined in Methods), we confirmed the final cohort stratification of 22 NRDS cases and 28 controls. Consequently, the cohort stratification (22 NRDS cases vs. 28 controls) remained identical to the initial analysis.

The study cohort comprised 50 GDM patients with mean maternal age 28.6 $\pm$ 3.4 years and mean gestational age 35.2 $\pm$ 2.1 weeks. As detailed in Table 1, glucose records for 4 controls were completed via multiple imputation, enabling full cohort analysis (n = 50). NRDS was diagnosed in 44.0% (22/50) of neonates, revealing a striking dichotomy in outcomes stratified by glycemic control. The cohort comprised 18 patients (36.0%) with suboptimal glycemic control and 32 (64.0%) with optimal control. Glycemic dysregulation manifested in both elevated uterine artery PI values (median 1.61 [IQR 1.50–1.75] vs. 1.21 [1.05–1.30], U = 63.5, p 0.001) and a 2.57-fold increased proportion of NRDS cases (suboptimal: 72.2% [13/18] vs. optimal: 28.1% [9/32]; p = 0.004, which remained significant after Bonferroni correction for the three primary group comparisons) among the 22 NRDS cases and 28 controls (Table 2), establishing hyperglycemia as the primary driver of the metabolic-vascular-pulmonary cascade [3]. The stratification of PI values by glycemic control status (suboptimal vs. optimal) in Fig. 2 mechanistically links maternal hyperglycemia to impaired placental perfusion, corroborating the 2.57-fold higher proportion of NRDS cases in suboptimally controlled pregnancies.

Fig. 2.

Uterine artery pulsatility index (UtA-PI) distribution stratified by NRDS status and glycemic control. (A) Significantly elevated UtA-PI in neonates with NRDS (n = 22) vs. controls (n = 28) (p 0.001, Mann-Whitney U test). (B) Metabolic-vascular interaction: UtA-PI gradient across glycemic control subgroups. Optimal control groups (Controls: n = 23; NRDS: n = 9) show lower PI than suboptimal groups (Controls: n = 5; NRDS: n = 13), demonstrating synergistic effects of hyperglycemia and NRDS on placental vascular resistance. NRDS, neonatal respiratory distress syndrome; PI, pulsatility index.

Uterine artery PI was significantly elevated in the NRDS group compared to controls (median 1.52 vs. 1.16, Table 1; U = 155.0, p 0.001). This significance held after strict Bonferroni correction for the three primary group comparisons (maternal age, gestational age, and PI) (p 0.001), with a large effect size (Cohen’s d = 1.59; 95% CI: 1.12–2.06) consistent with a priori assumptions. This establishes its role as a hemodynamic biomarker. Crucially, the stratification by glycemic control revealed a strong metabolic-vascular interaction (Table 2). The suboptimally managed subgroup manifested a significantly higher median PI than the target-achievers (1.61 [IQR 1.50–1.75] vs. 1.21 [IQR 1.05–1.30], Table 2; U = 63.5, p 0.001), mechanistically linking metabolic dysregulation to impaired placental perfusion. Significantly elevated uterine artery PI in the NRDS cohort vs. controls (median 1.52 vs. 1.16; p 0.001) is visually confirmed by the distributional disparity in Fig. 2, underscoring its role as a hemodynamic biomarker of placental insufficiency.

Post hoc power analysis confirmed that with the observed PI effect size (r = 0.62, Cohen’s d = 1.59) and n = 50, the achieved statistical power exceeded 99.0% (GPower 3.1), substantially surpassing the initial 80.0% target.

The AI-derived pulmonary texture signatures identified NRDS cases with 86.4% sensitivity (19/22) and 78.6% specificity (22/28). Model-clinical diagnosis concordance was substantial (κ = 0.67, 95% CI: 0.61–0.89), validating algorithmic reliability. The standalone AI classification achieved an overall accuracy of 82.0%. The corresponding confusion matrix, delineating true positives, false negatives, false positives, and true negatives, is provided in Fig. 3.

Fig. 3.

Confusion matrix evaluating AI model performance for neonatal respiratory distress syndrome (NRDS) prediction. The matrix compares actual clinical diagnoses (rows) against model predictions (columns) in the test cohort (n = 50 neonates), demonstrating 19 true positives (NRDS correctly predicted), 22 true negatives (controls correctly identified), 3 false negatives (NRDS cases missed), and 6 false positives (controls misclassified as NRDS). Key performance metrics derived include: sensitivity = 19/(19+3) = 86.4% (ability to detect NRDS cases), specificity = 22/(22+6) = 78.6% (ability to exclude controls), and overall accuracy = (19+22)/50 = 82.0%. The model shows stronger precision in NRDS identification than control group discrimination. TP, true positives; FN, False Negative; FP, False Positive;TN, True Negative.

The composite diagnostic criterion (concurrent AI high-risk classification and PI $\ge$1.28) demonstrated high clinical utility, achieving 92.0% overall accuracy (46/50) in the full cohort, with the complete breakdown of diagnostic performance metrics provided in Table 3. The decision curve analysis demonstrated a peak net benefit of 0.35 at the 20% threshold probability (Fig. 4), indicating optimal clinical utility when intervening for patients with $\ge$20.0% predicted risk.

Fig. 4.

Decision curve analysis (DCA) evaluating the clinical utility of predictive models for neonatal respiratory distress syndrome (NRDS). The net benefit (y-axis) of intervention strategies derived from the combined model (black), AI texture model only (red), UtA-PI model only (blue), treat-all approach (green), and treat-none strategy (purple) is plotted against threshold probabilities ranging from 10.0% to 50.0% (x-axis). The integrated (combined) model demonstrates superior net benefit compared to both standalone models across the entire threshold range (10.0%–50.0%). Crucially, at clinically actionable intervention thresholds exceeding 15.0%, the net benefit of the combined model is significantly higher than that of the best unimodal alternative (either AI or PI model). The combined model achieves a peak net benefit of 0.35 at the 20.0% threshold probability.

Decision curve analysis (DCA) quantified the net benefit profile across therapeutically relevant threshold probabilities (10.0%–50.0%). The integrated model consistently demonstrated superior net benefit compared to both standalone AI and standalone PI prediction methods across this range. Crucially, at clinically actionable intervention thresholds exceeding 15.0%, the net benefit of the combined model was significantly higher than that of the best unimodal alternative. Fig. 4 graphically validates these results, illustrating the net benefit superiority of the combined model across threshold probabilities (10.0%–50.0%), with peak net benefit of 0.35 at 20.0% threshold. This analysis confirms the enhanced clinical utility of the integrated diagnostic paradigm for guiding perinatal management decisions in GDM pregnancies at risk for NRDS.

Several methodological limitations warrant consideration. The single-center design and modest cohort size (n = 50) constrain immediate generalizability and necessitate external validation in larger, multiethnic populations [27]. Furthermore, potential confounding factors were not fully adjusted for in our analysis. These include not only established variables such as maternal body mass index [28], insulin therapy intensity, and fetal growth trajectories, but also the specific indications for elective premature delivery. Although major comorbidities were excluded, residual confounding from unreported pathologies associated with early delivery could independently influence NRDS risk. Additionally, the explanation of false-negative cases via a “metabolic threshold” remains speculative due to a lack of quantitative metabolic profiling, such as glycated hemoglobin (HbA1c), glucose variability, or insulin dosage, which limits our ability to precisely define the level of dysregulation in misclassified cases. Future studies should incorporate detailed metabolic, clinical, and delivery indication data to improve model specificity, explanatory power, and population-specific risk calibration [29].