PHYSIOLOGICALLY BASED FEATURES RELATED TO VOCAL HYPERFUNCTION: FROM LABORATORY TO AMBULATORY DATA

Abstract

The following thesis proposal describes a framework for the analysis of signal-based features related to vocal pathologies, namely phonotraumatic vocal hyperfunction (PVH) and non-phonotraumatic vocal hyperfunction (NPVH), using an accelerometer attached to the neck-skin in an ambulatory setting. The first stage consists of extracting physiologically relevant features that are associated with PVH on a daily basis. A clinical set-up (In Lab) that captures key components of vocal function, such as acoustics (microphone) and aerodynamics (oral air ow) from a reading passage, provides a set of model parameters to characterize vocal function. An impedance-based inverse filtering (IBIF) technique is used to estimate glottal airflow and related features from the accelerometer signal and to obtain the same features for the ambulatory data (In Field). An in-depth analysis of IBIF aerodynamic measures is done in the context of machine learning classifiers. Subsequently, an adaptive version of the IBIF filter (i.e., Kalman smoother) is proposed in order to estimate the airflow signal, incorporating modeling and observation noise. The Kalman smoother is compared to the original IBIF filter with In Lab and In Field data within a classification task to determine the efficiency and relevance of both approaches. Additional efforts are presented to provide insights on the capabilities of machine learning tools to be used on PVH and NPVH patients when compared to their matched-controls. First, a case study with 4 pairs of PVH and controls is used to determine how the variability of the IBIF parameters can affect the classification performance with In Lab data. Later, classical machine learning algorithms are used to investigate the nuances in the classification of NPVH subjects vs. controls, while a final effort explores the use of wavelets with deep learning to separate Pre vs Post therapy in NPVH patients. The main contributions of this thesis are: 1) to develop a machine learning framework for the analysis and classification of neck-surface acceleration signals using aerodynamic features; 2) to propose an alternative filtering scheme to IBIF based on adaptive filtering; and 3) to support the first two contributions by exploring pilot studies on salient features for NPVH, IBIF parameter uncertainty, and therapy effects on NPVH. Discussions and conclusions are included in each chapter to interconnect the ambulatory analysis of glottal flow with machine learning, to establish the potential benefits and limitations of these approaches in clinical settings.