DALA-XGBoost: An AI-ML Framework for Cross-Domain Maternal-Fetal Risk Assessment

Abstract

In the delicate balance of life before birth, even small maternal health issues can pose serious risks to the fetus. Recent reports indicate that over 260,000 women lost their lives in 2023 due to preventable pregnancy complications, which indicates the urgent need for proactive and automated methods to detect pregnancy risks early. The lack of clinically annotated maternal data sets, along with the problem of integrating maternal and fetal health records, limits the effectiveness of existing prediction tools. To address this, we introduce a new cross-domain classification framework that combines Dual Autoencoder Latent Alignment (DALA) with Extreme Gradient Boosting (XGBoost) classifier. In this approach, separate autoencoders are trained independently on the maternal and fetal datasets to learn domain- specific latent representations, and then aligned in a common space through unsupervised learning. Labeled fetal embeddings are used to pseudo-label unlabeled maternal data through distance clustering, thereby effectively transferring risk knowledge across domains. These maternal embeddings are then used to train an XGBoost classifier to predict the fetal risk categories Normal, Suspect, and Pathological exclusively based on maternal - indicators. The model obtained an overall accuracy of 95.8% and F1 scores of 98%, 89%, and 91% for the three risk classes. SHapley Additive exPlanations (SHAP) are used to highlight the key maternal features that influence predictions, Principal Component Analysis (PCA) validates the latent alignment, and Local Interpre table Model-agnostic Explanations (LIME) explain the individual pathological predictions for clinical transparency. In summary, this research offers a scalable and potent solution for detecting fetal risks using only maternal data, thus enabling early intervention in low-resource environments where fetal monitoring is unavailable.

Ähnliche Arbeiten