Get started
Work

Federated Multi-Task Learning for Privacy-Preserving Parkinson’s Voice Analysis Under Non-IID Labels

Year: 2026
Type: book-chapter
Abstract: Collaborative voice-based analysis for Parkinson’s disease (PD) is limited by privacy regulations, decentralized data ownership, and heterogeneous clinical annotations across institutions. Although federated learning (FL) enables privacy-preserving training, most approaches assume homogeneous prediction tasks and degrade under non-independent and non-identically distributed (non-IID) data. In clinical voice analytics, institutions often collect different types of labels, such as diagnostic classification labels or continuous severity scores, resulting in heterogeneous label spaces that limit standard federated approaches. This paper proposes a federated multi-task learning framework for Parkinson’s disease voice analysis that supports heterogeneous label spaces under non-IID data distributions. A shared acoustic encoder with task-specific heads is collaboratively trained for disease classification and Unified Parkinson’s Disease Rating Scale (UPDRS) severity regression, enabling shared representation learning across institutions without sharing raw patient data. Experiments on aligned features from the UCI Parkinson’s and UCI Telemonitoring datasets are conducted within a simulated federated setting with Dirichlet-based non-IID partitioning. The model achieves a mean F1-score of 0.80 for classification with low inter-client variability (standard deviation 0.004) and a mean R2 of 0.415 for severity estimation, with improved explained variance over centralized single-task baselines in the regression task. The learned representation further supports exploratory cross-task inference across institutions with heterogeneous label spaces without requiring additional supervision.
DOI: 10.3233/atde260668
Language: English
Open Access status: diamond
OpenAlex ID: W7204435246
FWCI:
Cites:
Cited by:
Related to: 10