Advancing Parkinson's Disease Monitoring: Developing AI-Enhanced Models for Predicting Disease Severity from Speech Data

No Thumbnail Available
Badani, Disha Yogesh
Issue Date
MSc in Business Analytics
Dublin Business School
This thesis develops machine learning models to predict Parkinson's disease (PD) severity scores from speech recordings. 22 vocal features related to frequency, amplitude, timing, jitter, shimmer etc. are extracted from sustained vowels, words, and sentences in the PD data. After preprocessing, exploratory analysis provides insights into feature distributions and correlations with severity scores like the Movement Disorders Society Unified Parkinson's Disease Rating Scale (MDS-UPDRS). Rigorous feature selection identifies the most predictive subsets. Multiple sophisticated algorithms including random forests, support vector machines and neural networks are benchmarked and tuned using cross-validation to predict the MDS-UPDRS scores. A nonlinear support vector regressor with optimal features achieves high accuracy. Thorough model interpretation explains performance, identifies limitations guiding future improvements and characterizes clinical implementation requirements. Overall, the interpretable modeling approach accurately forecasts PD severity, enabling potential telemonitoring applications.