Applying Advanced Data Analytics on Pregnancy Complications to Predict Miscarriage With eXplainable AI

Abstract

Pregnancy complications in the early months of the family process can lead to miscarriage. Miscarriage does not occur due to only one reason; many factors are involved in causing miscarriage. Deep Learning (DL) can help healthcare providers by providing advanced analysis. In this study, we have proposed two novel ensemble models, Echo Dense Inception Blending (EDI-Blend) and Dense Reservoir Inception Modular Network (DRIM-Net), for miscarriage prediction. The dataset is balanced through the use of the hybrid balancing technique NearSMOTE. As the dataset is high-dimensional, we have used the Absolute Shrinkage and Selection Operator to select essential features from the dataset that significantly impact pregnancy complications. We validate the output of our proposed EDI-Blend and DRIM-Net models using 10-Fold Cross Validation. To determine the contribution of features for miscarriage prediction two eXplainable Artificial Intelligence techniques are applied to EDI-Blend and DRIM-Net: Interpretable Model-agnostic Explanations and SHapley Additive exPlanations. We have compared our proposed EDI-Blend model with base models, and the results show that EDI-Blend model performance is more efficient, with 0.732 accuracy, 0.721 recall, 0.732 F1-score, and 0.721 Receiver Operating Characteristic-Area Under the Curve (ROC-AUC). The DRIM-Net model is also compared with baseline models and achieves 0.768 F1-score, 0.764 precision, 0.769 accuracy, 0.769 recall, and 0.837 ROC-AUC.

Keywords

• Deep learning
• pregnancy complications
• miscarriage
• NearSMOT
• least absolute shrinkage and selection operator
• feature selection