Early diagnosis of Parkinson’s disease from multiple voice recordings by simultaneous sample and feature selection

Highlights

• We validate the hypothesis that multiple samples per subject might degrade the accuracy.

• To enhance the accuracy, two dimensional data selection approach has been proposed.

• The proposed method selects samples and features simultaneously, and outperforms the existing methods.

Abstract

Parkinson’s disease (PD) is a serious neurodegenerative disorder. It is reported that more than 90% of PD patients have voice impairments. Multiple types of voice recordings have been used for PD detection. Previous work indicates that the use of multiple types of samples per subject degenerates PD detection accuracy. In this paper, we validate it, and propose a two dimensional data selection method for sample and feature selection. The proposed method ranks features by using chi-square statistical model, searches optimal subset of the ranked features and iteratively selects samples. Experimental results show that the proposed method outperforms the state-of-the-art methods in terms of PD detection accuracy on multiple types of voice data.

Introduction

Parkinson’s disease (PD) is a neurodegenerative disorder of central nervous system. It is the second most common neurological disorder after Alzheimer’s disease (AD) (Benba, Jilbab, & Hammouch, 2016a). It targets elder people mostly after the age of 60 years (Van Den Eeden et al., 2003). PD causes diverse symptoms which include bradykinesia (slowness of movement), dysphonia (voice impairments), rigidity, tremor, and poor balance (Cunningham, Mason, Nugent, Moore, Finlay, Craig, 2011, Dastgheib, Lithgow, Moussavi, 2012, Rigas, Tzallas, Tsipouras, Bougia, Tripoliti, Baga, Fotiadis, Tsouli, Konitsiotis, 2012). However, PD detection based on voice data has drawn significant attention because 90% of People with Parkinsonism (PWP) suffer from voice impairments (Naranjo, Pérez, Martín, & Campos-Roca, 2017). Moreover, diagnosis of PD based on voice signals is considered to be an early detection of the disease (Al-Fatlawi, Jabardi, Ling, 2016, Duffy, 2013, Sakar, Isenkul, Sakar, Sertbas, Gurgen, Delil, Apaydin, Kursun, 2013). These factors motivated the use of voice data for the PD diagnosis (Arora, Venkataraman, Zhan, Donohue, Biglan, Dorsey, Little, 2015, Hariharan, Polat, Sindhu, 2014, Orozco-Arroyave, Hönig, Arias-Londoño, Vargas-Bonilla, Daqrouq, Skodda, Rusz, Nöth, 2016, Orozco-Arroyave, Belalcazar-Bolanos, Arias-Londoño, Vargas-Bonilla, Skodda, Rusz, Daqrouq, Hönig, Nöth, 2015, Upadhya, Cheeran, Nirmal, 2018, Wu, Zhang, Lu, Guo, 2018). At early stages of PD, there are potential abnormalities in voice that might not be perceptible to listeners, but they can be evaluated by performing acoustic analysis on voice signals (Harel, Cannizzaro, Cohen, Reilly, & Snyder, 2004). Thus, there is a need of development of an expert system based on machine learning that can efficiently perform the acoustic analysis of voice data in order to discriminate between PWP and healthy subjects.

In recent years different researchers have proposed different non-invasive methods to detect PD using acoustic analysis of voice signals (Benba, Jilbab, Hammouch, 2016b, Das, 2010, Gürüler, 2017, Little, McSharry, Hunter, Spielman, Ramig, et al., 2009, Naranjo, Pérez, Campos-Roca, Martín, 2016, Naranjo, Pérez, Martín, 2017, Sakar, Isenkul, Sakar, Sertbas, Gurgen, Delil, Apaydin, Kursun, 2013, Tsanas, Little, McSharry, Spielman, Ramig, 2012). Sarkar et al. collected and analyzed multiple types of voice recordings from 40 subjects out of which 20 were healthy subjects and 20 were PWP (Sakar et al., 2013). They used support vector machine (SVM) and k-nearest neighbour (KNN) models and achieved mean accuracy of 55% using leave-one-subject-out (LOSO) cross validation (CV). To enhance the classification accuracy, different feature selection algorithms have been proposed (Benba, Jilbab, Hammouch, 2016a, Benba, Jilbab, Hammouch, 2016c, Cantürk, Karabiber, 2016, Gürüler, 2017, Khorasani, Daliri, 2014, Li, Zhang, Jia, Wang, Zhang, Xie, 2017, Ozcift, 2012, Parisi, RaviChandran, Manaog, 2018). For example, Canturk et al. used four feature selection algorithms and six different classifiers to enhance the classification accuracy and achieved accuracy of 57.5% for LOSO CV and 68.94% for 10 fold CV (Cantürk & Karabiber, 2016). Li et al. used hybrid feature learning and SVM for classification and achieved accuracy of 82.5% (Li et al., 2017). Benba et al. used mel frequence cepstral coefficients (MFCCs) for features extraction and SVM for classification (Benba et al., 2016a). They achieved classification accuracy of 82.5% for LOSO CV. Furthermore, Benba et al. used only vowel samples, a subset of human factor cepstral coefficients (HFCCs) features and achieved 87.5% accuracy for LOSO CV (Benba, Jilbab, & Hammouch, 2017), which indicates that some irrelevant samples may not help but even degenerate the detection accuracy.

In contrast to feature selection from multiple types of samples, our proposed method selects samples before feature selection. We validate the hypothesis that irrelevant samples that provide irrelevant patterns might degrade the PD detection accuracy of a predictive model. Therefore, we propose a novel simultaneous samples, features and hyper-parameters selection (SSFH) approach.

The rest of the paper is organized as follows; In Section 2, dataset and the proposed method are elaborated. In Section 3, validation scheme and evaluation metrics are discussed. While Section 4 is about experimental results. The last section is about conclusion.