Early diagnosis of Parkinson’s disease from multiple voice recordings by simultaneous sample and feature selection
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.
Section snippets
Dataset description
The multiple types of voice recordings dataset used in this study was collected and used by Sakar et al. (2013). The dataset contains voice recordings of 40 subjects, i.e., 20 PD patients and 20 healthy subjects. Recordings from each subject were performed by a Trust MC-1500 microphone with a frequency range between 50 Hz and 13 kHz. The microphone was placed at a distance of 15 cm from the subjects. It was set to 96 kHz and 30 dB. Twenty six samples were recorded from each subject. The first
Validation schemes for multiple samples per subject data and the problem of subject overlap
To evaluate the performance of a machine learning model, different validation schemes are used. The most commonly used validation schemes are hold-out validation, leave-one-out (LOO) and k-fold cross validation (CV). But these conventional validation schemes cannot be used with datasets having more than one samples per subject. Because, these validation schemes will introduce an artificial overlap of the same subject in training and testing datasets. To solve this problem, Sarkar et al.
Experimental results and discussion
In this section, three groups of numerical experiments are performed for comparison, i.e., all features and samples are used, noisy features are eliminated using χ2 statistical model and two dimensional data selection is performed with samples and features selected simultaneously. Experimental results show the performance improvement using the proposed two dimensional data selection approach. Moreover, the first three experiments are performed on the training database. And to validate the
Conclusion
In this paper, we have developed an expert system based on feature and sample selection for PD detection problem. It was pointed out that like irrelevant features, irrelevant samples also degrade the PD detection accuracy of the predictive model. Hence, a two dimensional data selection approach was proposed to simultaneously select optimal samples and optimal features. The proposed method achieved classification accuracy of 97.5% for LOSO CV on training database and 100% using testing database.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
CRediT authorship contribution statement
Liaqat Ali: Conceptualization, Formal analysis, Methodology, Software, Validation, Writing - original draft, Writing - review & editing. Ce Zhu: Investigation, Software, Resources, Supervision, Writing - review & editing. Mingyi Zhou: Formal analysis, Methodology, Validation, Visualization, Writing - original draft. Yipeng Liu: Conceptualization, Investigation, Resources, Supervision, Writing - review & editing.
Acknowledgement
This research is supported by National Natural Science Foundation of China (NSFC, No. 61602091, No. 61571102) and Sichuan Science and Technology Program (No. 2019YFH0008, No. 2018JY0035).
References (41)
- et al.
Detecting and monitoring the symptoms of Parkinson’s disease using smartphones: A pilot study
Parkinsonism & Related Disorders
(2015) - et al.
Using human factor cepstral coefficient on multiple types of voice recordings for detecting patients with Parkinson’s disease
IRBM
(2017) A comparison of multiple classification methods for diagnosis of Parkinson disease
Expert Systems with Applications
(2010)- et al.
Acoustic characteristics of Parkinsonian speech: Apotential biomarker of early disease progression and treatment
Journal of Neurolinguistics
(2004) - et al.
A new hybrid intelligent system for accurate detection of Parkinson’s disease
Computer Methods and Programs in Biomedicine
(2014) - et al.
Simultaneous learning of speech feature and segment for classification of parkinson disease
e-health networking, applications and services (Healthcom), 2017 IEEE 19th international conference on
(2017) - et al.
Addressing voice recording replications for Parkinson disease detection
Expert Systems with Applications
(2016) - et al.
A two-stage variable selection and classification approach for Parkinson disease detection by using voice recording replications
Computer Methods and Programs in Biomedicine
(2017) - et al.
Feature-driven machine learning to improve early diagnosis of Parkinson’s disease
Expert Systems with Applications
(2018) - et al.
Thomson multitaper MFCC and PLP voice features for early detection of Parkinson disease
Biomedical Signal Processing and Control
(2018)
Learning acoustic features to detect Parkinson disease
Neurocomputing
Classification of parkinson disease utilizing multi-edit nearest-neighbor and ensemble learning algorithms with speech samples
Biomedical Engineering Online
Tensor rank learning in CP decomposition via convolutional neural network
Signal Processing: Image Communication
Efficient diagnosis system for Parkinson’s disease using deep belief network
Evolutionary computation (CEC), 2016 IEEE congress on
A multiple-classifier framework for Parkinson disease detection based on various vocal tests
International Journal of Telemedicine and Applications
Analysis of multiple types of voice recordings in cepstral domain using MFCC for discriminating between patients with Parkinson disease and healthy people
International Journal of Speech Technology
Discriminating between patients with Parkinson and neurological diseases using cepstral analysis
IEEE Transactions on Neural Systems and Rehabilitation Engineering
Voice assessments for detecting patients with Parkinson diseases using PCA and NPCA
International Journal of Speech Technology
A machine learning system for the diagnosis of parkinson disease from speech signals and its application to multiple speech signal types
Arabian Journal for Science and Engineering
Cited by (111)
Optimal tuning of support vector machines and k-NN algorithm by using Bayesian optimization for newborn cry signal diagnosis based on audio signal processing features
2023, Chaos, Solitons and FractalsCitation Excerpt :Finally, the Chi-square test is employed as statistical filter to identify the most significant patterns from each set of acoustic features separately to faster information processing by each optimized classifier and improve its accuracy. It was found to be effective in identification of significant patterns with application to Parkinson's disease diagnosis [36], gene selection [37], and schizophrenia identification [38]. To sum up, the contributions of the current study are as follows:
Parkinson Disease Prediction Using CNN-LSTM Model from Voice Signal
2024, SN Computer ScienceAn evolutionary feature selection method based on probability-based initialized particle swarm optimization
2024, International Journal of Machine Learning and Cybernetics