Longitudinal voice biomarker trajectory modelling for Parkinson's disease severity: domain-adaptive transfer learning on mPower real-world smartphone data - Report - MDSpire
Advertisement
Longitudinal voice biomarker trajectory modelling for Parkinson's disease severity: domain-adaptive transfer learning on mPower real-world smartphone data
Clinical Report: Modeling Longitudinal Voice Biomarkers in Parkinson's Disease
Overview
This study introduces a domain-adaptive transformer model, DAT-PD, which predicts continuous severity trajectories in Parkinson's disease (PD) using smartphone voice recordings. The model demonstrates significant accuracy improvements over existing methods, achieving a mean absolute error of 2.74 MDS-UPDRS units.
Background
Parkinson's disease affects approximately 10 million people globally, with speech and voice changes impacting up to 90% of patients. Accurate and frequent assessment of motor symptom severity is crucial for effective management, yet traditional methods are often limited by the need for in-person evaluations. This study addresses the gap in longitudinal monitoring of PD severity through voice biomarkers.
Data Highlights
Metric
Value
95% CI
Mean Absolute Error (MAE)
2.74
2.44–3.01
Root Mean Squared Error (RMSE)
3.61
3.18–4.04
R²
0.93
0.91–0.95
Key Findings
DAT-PD achieved a mean absolute error of 2.74 MDS-UPDRS units, outperforming six baseline models.
eGeMAPS features significantly reduced MAE from 5.21 to 2.74 compared to MFCC-only representations.
SHAP-based analysis identified MFCC-2, Shimmer (APQ5), and Jitter as key voice biomarkers.
The model maintained robustness under low signal-to-noise ratios, supporting real-world application.
Domain-adaptive modeling effectively addresses variability from different smartphone devices and environments.
Clinical Implications
The findings indicate that DAT-PD can serve as a non-invasive tool for continuous monitoring of PD severity using smartphone recordings.
Conclusion
The development of DAT-PD represents a significant advancement in the use of voice biomarkers for monitoring Parkinson's disease severity, with implications for improved patient care and management.
