Integrating large language models and expert knowledge to identify risk factors and predict mode of delivery with machine learning.

Abstract

Background: Accurately predicting the mode of delivery is a significant challenge in obstetrics, with implications for maternal healthcare planning, resource allocation, and personalized patient counseling. While machine learning offers promising predictive capabilities, the selection of relevant clinical and socioeconomic risk factors from vast possibilities remains a critical and nontrivial step that directly affects model performance and clinical utility.

Objectives: This study aimed to (1) predict the mode of delivery (vaginal vs. cesarean) using machine learning algorithms and (2) identify the key risk factors influencing delivery mode prediction.

Methods: A retrospective study was conducted on eligible pregnant women who delivered at Bent-Al-Huda Hospital in Bojnourd, Iran. A hybrid feature identification approach was employed: large language models (LLMs) initially proposed 63 potential risk factors. Through iterative refinement involving domain experts and assessment of data availability and frequency, this list was consolidated to 35 final variables for modeling. Machine learning algorithms, including multilayer perceptron (MLP), decision tree, naïve Bayes, and support vector machine (SVM), were applied to build predictive models.

Results: Among all models, SVM showed the best performance using all features, achieving an accuracy of 94.8% and an area under the curve (AUC) of 0.94. Decision tree analysis identified previous cesarean section history, insurance type, and fetal presentation as the three most important predictive factors. After excluding 40 women with prior cesarean delivery, SVM maintained robust performance with accuracy of 94.3% and an AUC of 0.92. Using antenatal features only, MLP achieved the highest accuracy before and after excluding women with previous cesarean section (92.9% and 92.2%), while SVM also showed the highest discriminative ability (accuracy 92.4%, AUC 0.90 and accuracy 91.3%, AUC 0.87, respectively).

Conclusion: This study demonstrates that combining expert knowledge with LLMs enables effective risk factor identification and that careful feature selection allows the proposed model to predict delivery type with high accuracy.