Enhancing Obstetric Decision-Making With AI: A Systematic Review of AI Models for Predicting Mode of Delivery.

Accurate prediction of the mode of delivery is critical for optimizing maternal and neonatal outcomes and reducing unnecessary cesarean sections. In recent years, AI has emerged as a promising tool for enhancing obstetric decision-making. This systematic review aimed to evaluate and synthesize existing evidence on AI models developed for predicting the mode of delivery, comparing their performance and clinical applicability across diverse settings. A comprehensive literature search was conducted to identify studies that developed and/or validated AI-based predictive models for mode of delivery outcomes, including vaginal birth after cesarean, emergent cesarean section during labor, and spontaneous vaginal delivery failure. Seventeen studies meeting inclusion criteria were analyzed, encompassing various AI models such as Random Forest, Gradient Boosting, XGBoost, CatBoost, support vector machines, neural networks, QLattice, and ensemble methods. Key study characteristics, input variables, model performance metrics, validation methods, and findings were systematically extracted and compared. The included studies, conducted across multiple countries and healthcare settings, demonstrated generally good to excellent predictive performance, with area under the curve values. Real-time intrapartum data significantly enhanced model accuracy in several studies. Ensemble models and advanced machine learning techniques outperformed traditional logistic regression in many cases, although simpler models remained competitive when interpretability was prioritized. Common predictive variables included maternal age, parity, BMI, previous cesarean, sonographic findings, and cervical examination data. Model transparency and external validation were highlighted as critical considerations for clinical translation. AI models show substantial potential for improving the prediction of the mode of delivery and supporting obstetric decision-making. Ensemble and real-time dynamic models demonstrated the highest performance. However, challenges remain regarding external validation, model interpretability, and integration into clinical practice.