Objective: To evaluate whether acoustic features from voice recordings can distinguish Parkinson’s disease and related parkinsonian disorders from healthy controls.

Background: Speech production depends on coordinated respiratory, laryngeal, and articulatory control, all of which may be impaired in parkinsonian disorders. Prior studies suggest that acoustic abnormalities such as increased perturbation and reduced phonatory stability may serve as noninvasive digital biomarkers. Because voice samples can be collected remotely and at low cost, they may support screening where specialist evaluation is difficult to obtain.

Method: De-identified publicly available voice datasets were assembled for sustained phonation, isolated vowels, and connected speech, including read passages and spontaneous dialogue. Voice cohorts included 83 subjects in the four-class vowel dataset, 81 in sustained phonation Parkinson’s disease versus healthy control, 44 in vowel Parkinson’s disease versus healthy control, and 37 in text-based Parkinson’s disease versus healthy control. Recordings shorter than 1.0 second were excluded from vowel analyses to reduce feature-extraction artifacts. Thirty-six acoustic features were extracted per recording, including jitter, shimmer, harmonics-to-noise ratio, and summary spectral and cepstral statistics. Classification models used support vector machines with radial basis function kernels and standardized inputs. Hyperparameters were selected using five-fold cross-validation. Binary classifiers were trained for Parkinson’s disease versus healthy controls, and a four-class classifier was trained to distinguish healthy controls, Parkinson’s disease, progressive supranuclear palsy, and multiple system atrophy.

Results: Binary classification achieved holdout accuracy of 0.909 for sustained phonation and 0.795 for vowels. For connected speech, accuracy was 0.733 at the default threshold. Lowering the decision threshold to 0.30 increased Parkinson’s disease recall to 1.000 with overall accuracy of 0.867. In four-class vowel classification, holdout accuracy reached 0.852 after duration filtering.

Conclusion: Voice-derived acoustic features can support practical screening for Parkinson’s disease and related parkinsonian disorders. Because voice collection is rapid and scalable, these models may be useful in remote or resource-limited settings.

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/voice-based-screening-of-parkinsonian-disorders-using-acoustic-feature-analysis-and-machine-learning/