Signal Recognition for Parkinson's Disease Diagnosis Based on Multi-Source Deep Domain Generalization

Abstract:
Parkinson's disease (PD) is an incurable neurodegenerative disorder. Dysarthria, stemming from PD, in speech offers an accessible and non-invasive diagnostic indicator. However, PD speech data, with its limited sample size and high aliasing, presents challenges for machine learning. Thought multi-source deep transfer learning methods show promise in PD speech recognition, they typically rely on labelled target domain data. To overcome this limitation, our study proposes an unsupervised transfer learning approach, namely multi-source deep domain generalization (MDDG). MDDG comprises four key modules: feature extraction, inter-domain and classes adversarial learning, and classifier training solely on multisource datasets. Leveraging adversarial networks, MDDG extracts invariant features from source domains, minimizing both inter-domain distribution discrepancies and intra-class difference while maximizing the distance between different classes from multi-source domains. Experimental results demonstrate the superiority of the MDDG over recent methodologies, where MDDG achieves remarkable metrics: highest accuracy (75.80%), precision (66.48%), and specificity (87.61%), all with the lowest standard error, underscoring the effectiveness and stability of MDDG offering valuable support to medical professionals in efficient PD diagnosis and monitoring.