Clinical Scorecard: Modeling Longitudinal Voice Biomarkers to Assess Severity in Parkinson's Disease: Utilizing Domain-Adaptive Transfer Learning with mPower Smartphone Data

At a Glance

Key Highlights

• DAT-PD achieved a mean absolute error (MAE) of 2.74 MDS-UPDRS units.
• eGeMAPS features significantly outperformed MFCC-only representations.
• Robustness evaluated under harsh acoustic conditions with SNR as low as 0 dB.
• SHAP-based explainability identified key voice biomarkers: MFCC-2, Shimmer (APQ5), and Jitter.
• Supports unsupervised home-based monitoring of PD severity.

Guideline-Based Recommendations

Diagnosis

• Utilize continuous MDS-UPDRS Part-II scores for assessing PD severity.

Management

• Implement DAT-PD for ongoing evaluation of voice-based severity trajectories.

Monitoring & Follow-up

• Employ smartphone voice recordings for real-time monitoring of PD symptoms.

Risks

• Consider variability in acoustic environments affecting voice data quality.

Patient & Prescribing Data

Patients with Parkinson's disease experiencing voice and speech changes.

Voice biomarkers can provide a non-invasive method for continuous assessment.

Clinical Best Practices

• Incorporate domain-adaptive modeling to account for acoustic variability.
• Utilize SHAP explainability for clinician-interpretable feature identification.
• Leverage smartphone technology for remote patient monitoring.

Related Resources & Content

• mPower Dataset