A comparative review of machine learning models and deep learning architectures for pregnancy progression prediction

Abstract

Monitoring the course of pregnancy is of paramount importance, enabling the identification of negative outcomes and clinically significant changes. Although parity status may contribute to information needs, status as first-time mothers is not considered as a causal determinant. This paper presents a structured scoping review comparing conventional and deep learning approaches to pregnancy-related prediction tasks. The reviewed machine learning models include logistic regression; random forest; Extreme Gradient Boosting (XGBoost); Recurrent Neural Networks (RNNs) and their gated variants, long short-term memory (LSTM) and Gated Recurrent Units (GRUs); Convolutional Neural Networks (CNNs); transformers; and hybrid architectures. The analysis employs six dimensions: performance; ability to encode temporal data; model interpretability; data requirements for training; computational costs; and clinical relevance. As primary studies report outcome-specific performance results, the findings are presented across outcome domains rather than aggregated rankings of model architectures. In a study on predicting preterm-birth outcomes, a transformer achieved a test-set Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.792, compared to 0.78 for logistic regression, 0.751 for random forest, and 0.717 for a Support Vector Machine (SVM). In predicting preeclampsia, a CNN achieved an AUC of 0.85 on holdout validation and 0.81 on an external test set, with an AUC of 0.98 for early-onset preeclampsia. CTGFusionNet, which employs a hybrid architecture of convolutional, recurrent, and attention layers applied to Cardiotocography (CTG) and clinical records, reported 97.85% accuracy for the task of fetal-distress classification. The findings highlight the complementary strengths and weaknesses of conventional and deep learning models for the analyzed tasks. The choice of an architecture should be informed by the target outcome, data type, volume, and quality, calibration and validation strategies, requirements for interpretability and fairness, and computational costs rather than a singular emphasis on performance.