A Flexible Replication-Based Classification Approach for Parkinson’s Disease Detection by Using Voice Recordings

Abstract

Detecting Parkinson’s disease (PD) by using a noninvasive low-cost tool based on acoustic features automatically extracted from voice recordings has become a topic of interest. A two-stage classification approach has been developed to differentiate PD subjects from healthy people by using acoustic features obtained from replicated voice recordings. The proposed hierarchical model has been specifically developed to handle replicated data and considers a dimensional reduction of the feature space as well as the use of mixtures of normal distributions to describe the latent variables in the second order of hierarchy. The approach has been applied to a database of acoustic features obtained from 40 PD subjects and 40 healthy controls, improving results compared to previous models.