Privacy-Preserving Federated Unsupervised Learning for Diabetes Risk Pattern Discovery Across Heterogeneous Tabular Health Datasets

Abstract

Diabetes risk is influenced by interacting metabolic, anthropometric, cardiovascular, and lifestyle factors. Centralized machine-learning pipelines can discover latent patient subgroups, but they require pooling records from multiple data holders, which creates privacy and governance concerns. This paper presents a federated unsupervised learning framework for diabetes-risk pattern discovery using heterogeneous tabular health datasets. Public Kaggle datasets were used to simulate a multi-client healthcare environment under non-IID client partitioning. Federated standardization, Federated PCA, Federated KMeans, and federated centroid-distance anomaly scoring were compared with centralized PCA, centralized KMeans, and centralized Isolation Forest. The experiments covered 14 federated runs across Pima, BRFSS, and Health & Lifestyle data families, using five simulated clients per run. Federated PCA preserved the dominant variance direction with an average PC1 cosine similarity of 0.999. Federated KMeans achieved a mean ARI of 0.756 and a median ARI of 0.868 relative to centralized KMeans, while the median NMI was 0.781. Communication overhead was low, with an average estimated cost of only 0.072 MB per dataset. The findings indicate that federated unsupervised learning can recover major diabetes-risk structures without centralizing raw patient records, although highly non-IID partitions can degrade cluster agreement.