Accurate detection of Parkinson's disease (PD) through speech analysis holds great promise for early diagnosis and improved patient management. However, developing robust machine learning models is challenging due to the decentralized nature of medical data and the substantial heterogeneity in multilingual PD speech datasets. Conventional federated learning (FL) methods struggle in these heterogeneous, non-independent and identically distributed (non-IID) environments, where differences in data distributions arise from variations in language, speech content, recording conditions, medical measurement techniques, and dataset sizes. To address these challenges, we propose FedOcw, an optimized FL framework designed to enhance cross-lingual knowledge transfer and improve convergence stability. Through extensive multilingual experiments, we demonstrate that FedOcw consistently outperforms traditional FL models by achieving superior diagnostic accuracy while ensuring adaptive and equitable weight distribution across clients. These findings highlight FedOcw as an effective FL solution for privacy-preserving, speech-based PD detection across diverse linguistic and institutional settings.
© 2025. The Author(s).