What question did this study set out to answer?

March 14, 2026Digital Health0 citationsOpen Access

AI-powered speech analysis for early prediction of Parkinson's disease:

RCRitu ChauhanCentre for Artificial Intelligence and Robotics USUtkarsh SharmaSeoul National University of Science and Technology DSDhananjay SinghPennsylvania State University

Key Points

The research aims to evaluate if speech-derived acoustic features can serve as reliable early biomarkers for Parkinson's disease.
Analyzed 16 dysphonia features from 5875 speech samples of 42 individuals with Parkinson's disease.
Performed exploratory data analysis to examine feature distributions and correlations.
Used principal component analysis to reduce multicollinearity and identify predictors for machine learning models.
Compared performance of ensemble, kernel-based, and linear regression models using standard metrics.
Ensemble models showed the highest predictive performance with random forest explaining 91% of variance in motor UPDRS.
Gradient boosting explained 90% of variance in motor UPDRS, indicating robust predictive ability.
Linear models performed poorly, consistently yielding R² < 0.12, suggesting they lacked effectiveness in capturing non-linear dysphonia patterns.

Abstract

Objective Parkinson's disease (PD) is a progressive neurodegenerative disorder in which early diagnosis remains difficult due to subtle and heterogeneous symptoms. Speech impairments, particularly hypokinetic dysarthria, often appear early and offer promise as non-invasive biomarkers for detection. This study investigates whether quantitative speech-derived acoustic features can serve as reliable, non-invasive biomarkers for early detection of PD by analysing dysphonia measures extracted from sustained phonation recordings. Method A two-pronged analytical framework was used. First, exploratory data analysis was performed on 16 dysphonia features from 5875 sustained phonation samples collected from 42 individuals with idiopathic PD to examine feature distributions, correlations, and redundancies. Second, principal component analysis was applied to address multicollinearity among vocal features, and the resulting components were used as predictors in multiple regression-based machine learning models. Ensemble, kernel-based, and linear models were compared using standard metrics. Result Ensemble models delivered the strongest predictive performance. Random forest explained 91% of variance ( R 2 = 0.910 for motor Unified Parkinson's Disease Rating Scale (UPDRS); 0.901 for total UPDRS), with root mean squared error (RMSE) = 2.39 and 8.33 and mean absolute error (MAE) = 1.85 and 6.95, respectively. Gradient boosting explained 90% of variance ( R 2 = 0.900 for motor and total UPDRS), with RMSE = 2.52 and 8.42, and MAE = 1.86 and 7.14. Linear models performed substantially worse, consistently yielding R 2 < 0.12, indicating limited ability to capture nonlinear patterns in dysphonia characteristics. Conclusion Speech-derived acoustic biomarkers, when paired with machine learning, especially ensemble methods, show strong potential for accurate, scalable, and cost-effective assessment of PD severity. These findings highlight the potential of speech-derived acoustic biomarkers, coupled with machine learning, as scalable, cost-effective, and objective tools for improving diagnostic precision and enabling earlier intervention in PD.

KI fragen

Bookmark

View Full Paper