Abstract: The procedure of external cephalic version (ECV) is an important option in the management of breech presentation. However, there is still a lack of effective methods to accurately predict the likelihood of ECV success on the basis of individual conditions. With the aim of better predicting the outcomes of ECV and subsequent delivery modes, this study developed neural network-based models. We conducted a retrospective cohort study of women with singleton pregnancies who underwent an ECV for breech presentation at a single, tertiary, university-affiliated hospital between January 2016 and September 2023. Data on the demographic characteristics, comprehensive preoperative ultrasound assessment, and conditions during the ECV procedure were extracted from the hospital’s electronic record system. A neural network algorithm was implemented to establish prediction models for the success or failure of ECV, as well as subsequent delivery modes. The performance of the models was improved by increasing the number of iterations. A total of 378 patients were retrospectively included, including 279 successful and 99 unsuccessful cases, showing an overall success rate of 73.8%. Univariate analysis revealed that gravidity, parity, systolic blood pressure, presence of uterine fibroids, and amniotic fluid index (AFI) indicated by pre-operative and intra-operative ultrasound were positively associated with ECV success, while thicker maternal abdominal walls and the use of anesthesia were correlated with failure. Multivariate analysis determined that parity, uterine fibroids, AFI, and the use of anesthesia were independent determinants of ECV outcomes. The samples were then divided into training and testing sets in a 1:1 ratio. By increasing the number of iterations, the true positive rates for successful and failed version both reached 100%. The overall accuracy was 82.5% for predicting ECV outcomes. A total of 291 samples with delivery records were included for the prediction of delivery modes. The neural network-based model achieved a predictive accuracy of 78.8% after 2000 iterations, with the true positive rates for vaginal delivery and cesarean section both reaching 100%. These algorithms output exact prediction probabilities instead of providing the outcomes alone. An external validation set was also employed to further confirm the algorithm’s performance. Collectively, prediction models formulated on the basis of neural network algorithms were developed to assess the ECV outcomes and subsequent delivery modes, with both models demonstrating favorable results and great performance. For women eligible for ECV, the application of these models can potentially assist in the crucial clinical decision making process.