Pregnancy-related complications continue to be an important concern in maternal healthcare, particularly in situations where early identification of high-risk pregnancies can support timely medical intervention. Machine learning has shown promising results in predicting maternal health risks from physiological and clinical parameters. However, most conventional machine-learning approaches depend on centralized healthcare datasets, requiring sensitive patient information to be transferred to a common location. Such centralized data management can introduce concerns related to privacy, security, data ownership, regulatory compliance, and institutional data-sharing policies.Federated Learning (FL) offers an alternative by allowing healthcare institutions to collaboratively train a machine-learning model while keeping patient records within their respective institutions. However, healthcare data rarely follow identical distributions across hospitals. Differences in patient demographics, geographical conditions, disease prevalence, clinical practices, and available healthcare facilities can lead to highly non-independent and identically distributed (non-IID) data. Under such conditions, a single global model may not provide equally effective predictions for all participating institutions.To address this limitation, this study proposes an Adaptive Personalized Federated Learning (APFL) framework for privacy-preserving pregnancy risk prediction. The proposed framework combines personalized local models, adaptive client aggregation, differential privacy, and secure aggregation. Participating healthcare institutions act as decentralized clients and train the prediction model using locally available pregnancy-related data. Instead of sharing patient records, protected model updates are transmitted to a coordinating server. The server adaptively combines these updates, while personalized model components allow individual institutions to retain knowledge specific to their local patient population.The proposed framework will be evaluated using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC-ROC). Additional evaluation will consider communication overhead and privacy-related performance. The study aims to demonstrate that adaptive personalization can improve the robustness of federated pregnancy-risk prediction when healthcare data are heterogeneous, while simultaneously reducing the exposure of sensitive patient information. The proposed framework can provide a foundation for developing scalable and trustworthy artificial intelligence solutions for decentralized maternal healthcare.

Keywords: Federated Learning, Pregnancy Risk Prediction, Maternal Healthcare, Non-IID Data, Privacy-Preserving Machine Learning