Objective:

To develop a domain-adaptive transformer model (DAT-PD) for predicting continuous Parkinson's disease (PD) severity trajectories from smartphone voice recordings, addressing the need for effective monitoring of disease progression.

Approach:

• Model Development: DAT-PD was developed using the mPower dataset, which includes 58,247 voice recordings from 5,800 participants, incorporating noise-aware acoustic preprocessing, eGeMAPS feature extraction, a domain-adaptive attention mechanism, and a longitudinal trajectory decoder.
• Training and Evaluation: The model was trained and evaluated using continuous MDS-UPDRS Part-II scores as the prediction target, with confounder-aware domain adaptation to address demographic imbalances, including age differences between the PD cohort and healthy controls.

Key Findings:

• DAT-PD achieved a mean absolute error (MAE) of 2.74 MDS-UPDRS units (95% CI: 2.44–3.01) and an R² of 0.93 (95% CI: 0.91–0.95) on the test set.
• eGeMAPS features significantly outperformed MFCC-only representations, reducing MAE from 5.21 to 2.74.
• SHAP-based explainability identified MFCC-2, Shimmer (APQ5), and Jitter as the most influential voice biomarkers.

Interpretation:

The performance of DAT-PD suggests that domain-adaptive longitudinal modeling can capture voice-based severity trajectories in PD.

Limitations:

• The study primarily focuses on the mPower dataset, which may limit generalizability to other populations or datasets.
• Robustness under harsh acoustic conditions was evaluated, but real-world variability may still pose challenges, and potential biases in the dataset should be considered.

Conclusion:

DAT-PD may serve as a non-invasive tool for continuous PD severity assessment, supporting home-based monitoring.