Abstract:

Objective To develop an early automated differential diagnosis between Parkinson's disease (PD) and multiple system atrophy-Parkinsonism (MSA-P) using a non-invasive combination of voice signal analysis and artificial intelligence.
Methods From July, 2023 to February, 2025, a total of 48 MSA-P patients and 76 PD patients with a course of less than five years were recruited from Beijing Tiantan Hospital, Capital Medical University. Voice features, such as glottal, phonatory, articulatory, prosodic, phonological and representation learning-based features were extracted from eleven voice tasks. A data-driven approach was used to identify the most discriminative features, which were utilized to construct diagnostic models using a variety of machine learning models. The diagnostic model with the strongest discriminative efficiency was selected.
Results The logistic regression model showed the best performance. For early-stage patients with a course less than two years, the diagnostic accuracy, precision and recall rate between PD and MSA-P were 92.5%, 95.9% and 92.2%, respectively. For all the patients with a course less than five years, the logistic regression model achieved an accuracy of 89.1%, a precision of 91.6%, and a recall rate of 92.4%. Even when features extracted from a single speech paradigm were used for analysis, the diagnostic accuracy could still reach 77.7%.
Conclusion Voice signals analysis is potential in the early differential diagnosis of PD and MSA-P.

Key words: Parkinson's disease, multiple system atrophy, voice analysis, machine learning, differential diagnosis, early diagnosis