Janani-Net: A Privacy-Preserving Federated Learning Framework For Multi-Center Maternal Health Risk Stratification

17 Nov

Authors: Shruthi K R, Ashmaav Sharma, Nitisha Shajee, Rishit M. Udikeri

Abstract: Preventable maternal mortality and severe obstetric complications remain major public health concerns, especially in regions where patient data are fragmented across disconnected clinical facilities. Although machine learning models can identify high-risk pregnancies from routine physiological measurements, pool-ing raw clinical records into a central server introduces serious privacy risks and regulatory barriers. In this study, we introduce Janani-Net, a federated learning framework designed for collaborative maternal health risk stratification across multiple hospitals without transferring patient-level records. Implemented with Flower and PyTorch using a three-tier hub-and-spoke topology, Janani-Net simulates collaboration across five clinical centers with differing patient risk distributions and uneven computing resources. We compare standard Fed-erated Averaging (FedAvg) and Federated Proximal (FedProx) optimization under controlled non-independent and identically distributed (non-IID) Dirichlet class partitions and variable local training budgets. Across multi-center maternal health records, Janani-Net closely matches the accuracy of a centralized baseline model and consistently outperforms isolated single-hospital models while keeping raw data local. We also analyze round-by-round convergence, sensitivity to class imbalance, and run-to-run variance when statistical skew coincides with uneven local training epochs. These findings show that proximal federated optimization offers a practical, privacy-compliant path for deploying obstetric decision support across diverse hospital networks.

DOI: https://doi.org/10.5281/zenodo.24309874