Robustness of deep-learning speech biomarkers to OS noise-cancellation variability in longitudinal Parkinson's Disease monitoring
9th Annual Digital Biomarkers in Clinical Trials Summit · Basel · Hosted in partnership with Roche.
Comparative evaluation of four classifier architectures under simulated post-OS-update conditions, using the Italian Parkinson's Voice and Speech Dataset. Across six scenarios of mid-study noise-cancellation algorithm change, our dual-pipeline classifier maintained sensitivity within an 11.3 percentage-point spread; three baseline architectures showed spreads of 42 – 46 percentage points. Reporting follows TRIPOD+AI.
Technical report · Selected poster · Available under MTA