Predicting preterm birth with privacy-preserving AI models: Federated learning and explainable AI

Preterm birth remains a significant public health challenge, closely associated with infant mortality and long-term morbidity. The complexity of its causes complicates accurate prediction. In this study, we present an AI-driven model designed to predict preterm birth, integrating federated learning (FL), deep learning (DL), and explainable artificial intelligence (XAI) to prioritize both data privacy and interpretability. We utilized a primary dataset of 58 electrohysterogram (EHG) recordings from pregnant women, each collected over 1000-second intervals, and applied the Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance. To rigorously assess generalizability, we performed external validation on the independent TPEHGDB dataset comprising 300 EHG recordings from a different institution and time period Our approach evaluated a range of models, from established machine learning algorithms like XGBoost, LightGBM, and CatBoost, to advanced frameworks such as a Transformer-based architecture and quantum convolutional neural networks (QCNN). By leveraging FL, we enabled secure, collaborative training across institutions while maintaining patient data confidentiality. Additionally, XAI techniques, particularly SHAP, were employed to elucidate the key risk factors influencing predictions, thereby enhancing clinical transparency. XGBoost and Transformer models achieved 96.17% and 94.94% accuracy on internal validation, respectively, and demonstrated robust generalization with 88.67% and 89.33% accuracy on external validation, maintaining clinically critical recall rates of 78.95% and 81.58% for preterm detection. Critically, federated learning introduced minimal performance degradation (more than 2%) compared to centralized training, validating privacy-preserving collaborative learning. Although QCNN showed promise as an innovative approach, its performance lagged slightly behind classical models on external data. This underscores the potential of our approach as a scalable, privacy-preserving, and interpretable tool for early detection of preterm birth, with demonstrated generalizability across independent clinical populations.