(a) Traditional ML Models: De Ramón Fernández A et al. [11] employed Support Vector Machine (SVM), Multilayer Perceptron (MLP), and Random Forest (RF) on a large retrospective cohort (n = 25,038), achieving $\ge$90% accuracy in distinguishing cesarean versus vaginal delivery and 86–87% accuracy for instrumental versus eutocic vaginal delivery. Notably, RF demonstrated marginal superiority (accuracy 91% vs. SVM 90%, MLP 90%), suggesting ensemble methods’ robustness in handling heterogeneous obstetric data.

Ullah Z et al. [2] compared five ML algorithms (DT, RF, AdaBoostM1, Bagging, k-NN) on a small enriched dataset ((synthetic minority oversampling technique) SMOTE-augmented n = 80). K-NN achieved the highest accuracy (84.38%), while AdaBoostM1 performed poorest (80.63%). Data augmentation improved model performance by 3–5%, highlighting its utility in addressing class imbalance.

(b) Deep Learning (DL) Innovations: Kuanar A et al. [12] pioneered Deep Neural Networks (DNN) adoption (n = 101), reporting exceptional training metrics (Area Under the ROC Curve (AUC) = 0.99, Kolmogorov-Smirnov statistic (KS) = 0.98) but limited external validity due to minimal sample size. Prediction error rates for cesarean and vaginal delivery were 0.02 and 0.00, respectively, though these results may reflect overfitting.

(a) Scale Disparity: Large-scale studies [11] (n = 25,038) demonstrated stable performance (AUC 0.87–0.91), whereas small cohorts [2, 12] (n = 80–101) exhibited inflated metrics (AUC up to 0.99), likely due to limited generalizability.

(b) Feature Engineering: Maternal age, Body Mass Index (BMI), and parity (Static Parameters) dominated input features across studies [2, 11]. None incorporated real-time intrapartum progression metrics (e.g., cervical dilation rate), constraining clinical utility [11].

(a) Internal Validation: All studies used random splits, with only De Ramón Fernández A et al. [11] achieving NOS = 9 through rigorous sensitivity analysis.

(b) Temporal/External Gaps: No study implemented prospective temporal validation, and inter-hospital transportability tests were absent.

(a) Traditional ML Models: Ferreira I et al. [13] developed a multivariable logistic regression (LR) model using retrospective data from singleton term pregnancies (n = 2434). The LR model achieved an areas under the receiver-operating-characteristics curve (AUROC) of 0.794, with Bishop score, maternal height, and inter-delivery interval identified as top predictors through SHAP analysis. Despite moderate discrimination, the model prioritized clinical interpretability over complex architectures. This study excluded real-time intrapartum parameters, relying solely on static admission features, which constrained its utility in dynamic labor management.

(b) Automated Machine Learning (AutoML) Innovations: Wong MS et al. [14] pioneered AutoML (Partometer) using real-time intrapartum data (n = 37,932). The model achieved 87.1% accuracy and AUC = 0.82 in predicting vaginal delivery within 4 hours of admission. Key dynamic predictors included cervical dilation rate and fetal head descent, underscoring the value of temporal feature integration. While AutoML streamlined model development, the “black-box” nature of feature selection reduced clinician interpretability, a critical barrier to adoption.

(a) Scale and Diversity: Large-Scale Dynamic Data [14]: The inclusion of 37,932 records with real-time monitoring metrics provided robust statistical power, though the cohort was limited to primiparous women in high-resource settings.

Static Data Limitations [13]: Despite a moderate sample size (n = 2434), reliance on retrospective, single-center data impeded generalizability to diverse obstetric populations.

(a) Internal Validation: Both studies employed cross-validation [13, 14], with Wong MS et al. [14] achieving superior discrimination (AUC = 0.82) due to dynamic feature inclusion.

(b) Temporal and External Gaps: No External Validation: Neither study tested models across institutions or regions, risking overfitting to local practice patterns.

Real-World Implementation: While Wong MS et al. [14] demonstrated the feasibility of real-time prediction, the absence of clinician-AI interaction protocols limited practical utility.

(a) Traditional ML Models: Meyer R et al. [18] employed XGBoost on a large cohort (n = 73,667), achieving AUC = 0.84 for unplanned CS prediction. Feature importance analysis identified maternal BMI and labor progression rate as top predictors, aligning with clinical intuition.

Islam MS et al. [19] proposed the Henry Gas Solubility Optimization-based Random Forest (HGSORF) algorithm (optimized RF), reporting 98.34% accuracy on a balanced dataset (n = 15,409). Explainable AI (XAI) analysis revealed placental insufficiency and uterine contractility patterns as critical drivers, enhancing model interpretability.

(b) Deep Learning (DL) and Hybrid Models: Fergus P et al. [16] applied deep learning classifiers to fetal heart rate (FHR) signals (n = 552), achieving 94% sensitivity and 91% specificity. AUC = 0.99 underscored DL’s potential but raised concerns about overfitting in small samples.

Nagayasu Y et al. [17] utilized the recursive-rule eXtraction (Re-RX) rule extraction method (n = 1513), yielding 81.9% accuracy and AUC = 0.71. While interpretable, the model’s lower discrimination highlighted trade-offs between simplicity and predictive power.

(c) Cross-Institutional Validation: Guedalia J et al. [15] tested model transportability between hospitals, finding performance drops ($\mathrm{\Delta }$AUC: –0.12) due to interfacility measurement variability. Adjustments to fetal head position metrics restored parity, emphasizing standardized feature protocols.

(a) Scale and Diversity: Large-Scale Cohorts [18, 19]: Studies with $>$15 k records demonstrated stable performance (AUC 0.84–0.98), whereas smaller datasets [16, 17] (n = 552–1513) exhibited inflated metrics (AUC up to 0.99).

(b) Dynamic Data Integration: Only Fergus P et al. [16] incorporated real-time FHR signals, while others relied on static admission parameters (e.g., parity, BMI) [17, 18, 19].

(a) Internal Validation: All studies used cross-validation [15, 16, 17, 18, 19], with Islam MS et al. [19] achieving the highest accuracy (98.34%) through Adaptive Synthetic Sampling Approach (ADASYN)-balanced data.

(b) External and Temporal Gaps

Limited Generalizability: Only Guedalia J et al. [15] addressed inter-hospital variability, revealing institutional bias as a critical barrier.

Real-Time Application: Despite high accuracy, no study has implemented prospective real-time prediction in clinical workflows.

(a) Traditional ML Models: Lindblad Wollmann C et al. [3] compared ML models (conditional RF, lasso regression) with existing clinical scores (Swedish cohort, n = 3116). All models achieved AUROC 0.61–0.69, with sensitivity $>$91% but specificity 22%, indicating strong ability to identify VBAC candidates but poor discrimination for repeat cesarean risks.

Meyer R et al. [20] implemented RF and XGBoost (n = 989), reporting AUC-PR 0.351 (RF) vs. 0.325 (MFMU model). The XGBoost model required only 8 variables, emphasizing parsimony but sacrificing nuanced risk stratification.

(b) Dynamic Risk Stratification: Lipschuetz M et al. [21] developed gradient boosting models using first-trimester and pre-labor data (n = 9888). The pre-labor model achieved AUC = 0.793, significantly outperforming the first-trimester model (AUC = 0.745). Risk stratification categorized 42.4% of women as low-risk (VBAC success 97.3%, demonstrating clinical utility for personalized counseling.

(a) Scale and Diversity: Large National Cohorts [3]: Population-based data (n = 3116) enhanced generalizability but lacked granular intrapartum metrics (e.g., cervical dilation trends).

Real-Time Data Gaps: All studies relied on retrospective, static parameters (e.g., prior vaginal delivery, maternal BMI), neglecting dynamic labor progression [20, 21].

Geographic Bias: Studies focused on high-income populations (Sweden [3], Israel [21]), limiting applicability to low-resource settings.

(a) Internal Validation: Lindblad Wollmann C et al. [3] used cross-validation but reported low specificity (22%), reducing clinical confidence in avoiding unnecessary cesareans.

Lipschuetz M et al. [21] demonstrated temporal validity with pre-labor data integration, aligning closer to clinical workflows.

(b) External and Practical Gaps: No Inter-Hospital Testing: Despite Meyer R et al. [20]’s multi-algorithm comparison, no study validated models across institutions, risking practice pattern overfitting.

Patient-Clinician Discordance: Meyer R et al. [20] noted 28% patient refusal of VBAC attempts due to anxiety, a psychological factor absent in AI frameworks.

AI is a transformative technology that aims to simulate, extend, and augment human intelligence through the development of advanced algorithms and data analysis techniques. It can concurrently handle and analyze a vast amount of clinical data. In facilitating the prediction of delivery methods, it can fully exploit the advantages of big data to enhance the accuracy and reliability of predictions. Moreover, AI enables individualized clinical decision-making. In the current era marked by the progressive improvement of electronic health records, AI can leverage its strengths to comprehensively analyze historical big data. Based on the specific circumstances of each pregnant woman and the current pregnancy examination data, it can offer more personalized guidance. Simultaneously, it can provide more objective support for the clinical diagnosis and treatment work of obstetricians, optimize delivery outcomes, and continuously elevate the health status of mothers and infants.