Exploring Feature Selection and Feature Transformation Techniques to Improve Telephone-based Biomedical Speech Signal Processing towards Parkinson's Assessment

Clinical decision support tools mining speech signals for Parkinson's Disease (PD) applications typically rely on relatively small numbers of participants, having collected data under highly controlled acoustic conditions. We recently reported on the Parkinson's Voice Initiative (PVI), a large international project leading to the collection of 19,000+ sustained vowel phonations (control and PD groups) across seven countries, where participants were self-selected and provided phonations over the standard telephone network. In this study, we explored sustained vowels in a balanced subset of the US-speaking cohort in PVI comprising 3000 participants (1500 PD and 1500 controls). The aim was to investigate feature selection and feature transformation techniques towards improving binary differentiation of controls and PD and obtaining new insights in a lower dimensional space. We acoustically characterized each sustained vowel /a/ phonation using 307 dysphonia measures which had previously been successfully employed in speech-PD applications. We explored five different feature selection and two manifold embedding techniques to project data into new feature spaces which might be more predictive of the binary outcome, and presented those into a Random Forest. We assessed the performance of the resulting models using internal 10-fold Cross-Validation (CV). We report classification accuracy of 67% and provide tentative interpretation by comparing the different feature subsets identified using different methods. Collectively, these findings provide new insights towards developing parsimonious feature subsets to facilitate the development of a large-scale tool for PD screening at minimal cost using telephone-based sustained vowels.