Automatically generated speech biomarkers from text reading differentiate Parkinson’s disease and multiple system atrophy

Background: Parkinson’s disease (PD) and multiple system atrophy (MSA) are neurodegenerative disorders with overlapping speech impairments but distinct pathophysiology. Automatically generated speech biomarkers may provide objective, and scalable markers to improve diagnostic accuracy. Objectives: To determine whether automatically generated speech-derived features can differentiate PD from MSA. Methods: Forty-three participants (22 PD, 21 MSA) completed a reading task, from which the ki:SB-M intelligibility score and 70 acoustic features were extracted. Group-level analyses and supervised machine learning algorithms were employed for classification. Results: Individuals with MSA showed lower intelligibility and significant differences in 16 acoustic features, including reduced loudness, diminished articulatory dynamics, less stable phonation, and greater utterance variability. The best speech-based classifier model achieved an AUC of 0.97, outperforming models using MDS-UPDRS scores alone (AUC = 0.56–0.85). Conclusions: Speech-derived features provide robust markers for differentiating PD from MSA. These findings highlight the potential of speech analysis to enhance early differential diagnosis.