Knowledge Discovery Improvement through a Federated Self-Organizing Map

Resumo

Healthcare privacy regulations prohibit the sharing of identifiable patient data across institutions, forcing hospitals to rely exclusively on local datasets for model development. This constraint results in limited-scope training data that typically yields models with poor generalizability. While traditional Federated Learning (FL) preserves data privacy by enabling hospitals to train local data and share only model updates (rather than raw data), this approach suffers from slow convergence since it incrementally updates models with new datasets. Also, FL favors larger datasets, which may marginalize small groups such as rare diseases. To address these limitations, we propose Federated Self-Organizing Map (FSOM), a topology-based approach that reduces bandwidth demands through less frequent data transmissions and boosts accuracy through structural data preservation. FSOM's topology-driven framework naturally accommodates small population groups in the analysis by maintaining their distinct data patterns during federated aggregation. This manuscript compares the performance of the traditional FL model with the proposed FSOM approach in uncovering hidden knowledge within the dataset. We evaluated the model using the Pima Indians Diabetes dataset.

Our results demonstrate that FSOM outperforms traditional FL, achieving a 2.55% increase in accuracy and a 4.73% improvement in macro F1-score. While accuracy reflects overall prediction correctness, the macro F1-score ensures balanced performance across all population subgroups—including minority classes. Notably, FSOM's improvements in macro F1-score demonstrate its particular value for marginalized demographic groups. By leveraging diverse data sources through federated learning, the global model inherently reduces bias toward majority populations that often dominates single-institution datasets.