Abstract

Analyses utilizing medical data from multiple healthcare institutions often face governance and privacy constraints, preventing individual facilities from sharing patient-level data externally. In addition to these constraints, when significant heterogeneity exists between facilities in terms of patient populations, clinical practices, and the prevalence of outcomes, further challenges arise in constructing reliable and generalizable models. These issues are particularly pronounced in facilities with limited sample sizes, where model estimation and performance evaluation may become unstable.
To address these challenges, this study adopts a personalized federated learning framework that enables collaborative training without compromising data privacy while capturing facility-specific heterogeneity. The proposed method adopts AUC maximization as the objective function for model learning, focusing on optimizing ranking-based discriminatory performance. As AUC is independent of diagnostic thresholds, score-based ranking avoids explicit threshold selection and provides relatively stable estimation even under class-imbalanced conditions.
Furthermore, we incorporate geographical proximity between facilities through a regularization term based on spatially clustered coefficient regression. This normalization encourages model coefficients to be similar among geographically proximate healthcare institutions, reflecting shared regional characteristics and healthcare environments. By leveraging information from neighboring facilities, the proposed method stabilizes model estimation and is expected to improve diagnostic discriminatory ability, particularly in facilities with small sample sizes. Discriminatory performance is quantified using the area under the receiver operating characteristic curve (AUC), a widely used threshold-independent measure in medical applications. Numerical experiments demonstrate the effectiveness of the proposed method.

Co-authors: Keisuke Hanada, Ke Wan, Kensuke Tanioka, and Toshio Shimokawa