Federated anomaly detection for neonatal health: Evaluating anomaly score aggregation functions for low birth weight and preterm birth detection in parous and nulliparous women
Ali Nawaz, Shehroz S. Khan, Amir Ahmad, Nadirah Ghenimi, Luai Awad Ahmed · Informatics in Medicine Unlocked · 2025
Supervised learning for neonatal risk classification suffers from severe class imbalance and severe data privacy constraints, limiting generalizability and cross-institutional deployment. We present a privacy-preserving, i.e., federated anomaly detection framework for early detection of neonatal risks, including low birth weight (LBW), very low birth weight (VLBW), extreme low birth weight (ELBW), and preterm birth (PTB) in parous and nulliparous women. Each participating client trains a local autoencoder on normal cases only, computes reconstruction-based anomaly scores, and sends updates to the server. Particularly, in score-level aggregation, clients transmit only anomaly scores, whereas in FedAvg, clients transmit model weights. In both cases, raw data remains local. We comparatively evaluate aggregation strategies for combining client anomaly signals: mean, median, maximum, minimum, and FedAvg. Across five-fold cross-validation, varying client counts (2–7), and both equal and unequal client data assignments, simple score-level aggregation by minimum achieved the highest number of wins on Area Under the Receiver Operating Characteristic Curve (AUC-ROC) and Area Under the Precision-Recall Curve (AUC-PR), outperforming FedAvg, which only combines model parameters. Our results indicate that in privacy-preserving neonatal analytics, robust score-level aggregation can be a stronger lever than weight averaging, especially under non-IID, imbalanced conditions. We discuss clinical implications for triage pathways and integration into multi-center maternal-fetal health networks. • First comparison of mean, median, min, and max aggregators with FedAvg for neonates. • Study of client heterogeneity via varying client counts and data distribution. • FedAvg extended with new aggregation in federated anomaly detection for neonates.