Process Systems Engineering Guided Machine Learning for Speech Disorder Screening in Children

Abstract

Abstract Auditory perceptual analysis (APA) is the primary method for clinical assessment of speech-language deficits, one of the most prevalent childhood disabilities. Due to multiple limitations of APA including being susceptible to intra- and inter-rater variabilities, automated methods such as Landmark (LM) analysis that quantify speech patterns for diagnosing speech disorders in children are developed. This work investigates the utilization of LMs for automatic speech disorder detection in children. Leveraging the similarities between disease detection in medical/clinical research and fault detection in process systems engineering (PSE), we propose to improve the detection of speech disorder in children via PSE principles. Specifically, the parsimony principle is followed for reducing feature and parameter spaces. Domain knowledge is utilized for generating a set of novel knowledge-based features to address the challenge of large within-class variations in LM measurements. A systematic study and comparison of different linear and nonlinear machine learning classification techniques are conducted to assess the effectiveness of the novel features in classifying speech disorder patients from normal speakers.