A Diadochokinesis-based expert system considering articulatory features of plosive consonants for early detection of Parkinson's disease

Abstract

Background and Objective: A new expert system is proposed to discriminate healthy people from people with Parkinson's Disease (PD) in early stages by using Diadochokinesis tests. Methods: The system is based on temporal and spectral features extracted from the Voice Onset Time (VOT) segments of /ka/ syllables, whose boundaries are delimited by a novel algorithm. For comparison purposes, the approach is applied also to /pa/ and /ta/ syllables. In order to develop and validate the system, a voice recording database composed of 27 individuals diagnosed with PD and 27 healthy controls has been collected. This database reflects an average disease stage of 1:85+ -0:55 according to Hoehn and Yahr scale. System design is based on feature extraction, feature selection and Support Vector Machine learning. Results: The novel VOT algorithm, based on a simple and computationally efficient approach, demonstrates accurate estimation of VOT boundaries on /ka/ syllables for both healthy and PD-affected speakers. The PD detection approach based on /k/ plosive consonant achieves the highest discrimination capability (92.2% using 10-fold cross-validation and 94.4% in the case of leave-one-out method) in comparison to the corresponding versions based on the other two plosives (/p/ and /t/). Conclusion: A high accuracy has been obtained on a database with a lower average disease stage than previous articulatory databases presented in the literature.