A recurrence plot-based approach for Parkinson’s disease identification

https://doi.org/10.1016/j.future.2018.11.054Get rights and content

Highlights

  • We proposed the usage of recurrence plots for Parkinson’s disease identification.

  • Recurrence plots provided more discriminative representations.

  • Significant improvement was accomplished using the proposed approach.

  • An average accuracy of over 87% was achieved.

Abstract

Parkinson’s disease (PD) is a neurodegenerative disease that affects millions of people worldwide, causing mental and mainly motor dysfunctions. The negative impact on the patient’s daily routine has moved the science in search of new techniques that can reduce its negative effects and also identify the disease in individuals. One of the main motor characteristics of PD is the hand tremor faced by patients, which turns out to be a crucial information to be used towards a computer-aided diagnosis. In this context, we make use of handwriting dynamics data acquired from individuals when submitted to some tasks that measure abilities related to writing skills. This work proposes the application of recurrence plots to map the signals onto the image domain, which are further used to feed a Convolutional Neural Network for learning proper information that can help the automatic identification of PD. The proposed approach was assessed in a public dataset under several scenarios that comprise different combinations of deep-based architectures, image resolutions, and training set sizes. Experimental results showed significant accuracy improvement compared to our previous work with an average accuracy of over 87%. Moreover, it was observed an improvement in accuracy concerning the classification of patients (i.e., mean recognition rates above to 90%). The promising results showed the potential of the proposed approach towards the automatic identification of Parkinson’s disease.

Introduction

First described by James Parkinson in 1817, Parkinson’s disease (PD) is a neurodegenerative disease affecting around 7 to 10 million people worldwide, being nearly 1 million people only in the United States of America. Also, according to the National Parkinson’s Foundation [1], the number of new cases diagnosed each year ranges from 50,000 to 60,000 individuals. PD is caused by the loss of a neurotransmitter called Dopamine [2], being a chronic, progressive and multi-lesion disease. The lack of Dopamine increases the malfunctioning of the brain affecting the communication among its cells. As a consequence, messages are not sent correctly, which may lead to cases of depression, sleep disturbances, dementia, depression, psychotic features, and autonomic dysfunction [3]. PD patients may also face motor dysfunctions, such as tremor, and an increase in the risk of falls and mobility disabilities due to alterations in gait and posture. It has a tremendous negative impact on the daily routine of PD patients and their families, thus reducing their quality of life [4], [5], [6].

Although non-lethal, life expectancy is shorter for PD patients compared with the general population. Its diagnose is usually performed using a clinical exam and by a neurologist with expertise in movement analysis. The current treatments for PD symptoms include therapy and drugs known as dopaminergic medications, being the Levodopa (L-dopa) the most widely used for such purpose. Another widely employed treatment is the Deep Brain Stimulation, which is a surgical procedure that delivers electrical pulses to brain cells to reduce the effects of the symptoms.

The science does not measure efforts to improve the quality of life regarding PD patients [7], [8], [9], [10]. Computer-aided detection and diagnosis systems based on image processing, neural networks, and other techniques have been widely applied in the pursuit of better results in both treatment and aid of diagnosis [11]. One of the motivations for the development of such tools comes from the fact that pattern recognition and other techniques may be able to detect subtle signs assigned to PD that might not be noticed by a human. Also, they may provide some relevant information to support the final diagnosis. Spadotto et al. [12], for instance, proposed the automatic PD identification using the Optimum-Path Forest (OPF) [13], [14] classifier. A later work of the same group proposed the selection of a subset of features through evolutionary-based techniques, which helped to improve PD identification accuracy [15], [16], [17].

Most works that address automatic PD identification cope with voice-based data. Procedures to identify voiced and unvoiced(silent) periods have been actively pursued to analyze continuous speech samples since most techniques that quantify periodicity and regularity in voice signals are applied in the voiced regions only [18]. Das [19] compared different classification methods (e.g., neural networks, decision trees, and regression) for the diagnosis of PD. The experiments were carried out in a dataset comprised of biomedical voice measurements collected from 31 people, being 23 diagnosed with PD and eight healthy individuals. The approach was evaluated under several methods, with the best results achieved by neural networks with an accuracy of around 92.9%.

However, the clinical diagnosis of PD can also be performed by the identification of some motor signs such as bradykinesia, tremor, rigidity, and even postural instability [3]. A few studies were dedicated to monitor, assist, and aid the detection of PD based on signals collected from accelerometers and other sensors placed in wearable devices and biometric pens, which can provide valuable information of any existing motor dysfunction. Recently, Mamun et al. [20] proposed a cloud-based framework to allow the diagnosis and monitoring of PD. The individual interacts with the doctor through mobile devices connected to the internet. Voice samples are recorded and have features extracted for the diagnosis using machine learning techniques.

Khan et al. [21] worked with different accelerometer data including those from PD patients performing daily activities to find an optimum sampling for human activity recognition. Sama et al. [22] proposed to detect the freezing of gait (FoG), another common symptom in PD patients, based on data collected by a single-waist accelerometer in home environments. The work aimed at the monitoring of FoG to apply real-time cueing strategies by selecting a subset of optimal features. Bächlin et al. [23] introduced a wearable assistant for PD patients that present FoG called “wearable health assistant”, which aims to reduce the number and length of their motor blocks, and thus increase their safety while walking by performing an online movement analysis and providing a rhythmic auditory signal that stimulates the patient to resume walking.

A work on accelerometer-based posture analysis in PD was proposed by Palmerini et al. [24]. A total of 175 measurements are computed from the accelerometer data, and a feature selection technique was applied to select the most discriminative set of measurements to characterize PD and healthy control (HC) subjects. The subset of measurements was used to quantify tremor, acceleration, and displacement of body sway.

Moreover, a significant number of works explored tremor,which is the most common disorder in PD patients. Rigas et al. [25] proposed an automated method to assess both resting and action tremors by estimating the type and severity of data acquired from accelerometers attached to specific positions at a patient’s body. The estimation is based on features extracted from the signals and Hidden Markov Models. Abdulhay et al. [26] used tremor along with gait and machine learning techniques for the diagnosis of PD. They extracted features from signals recorded by sensors placed underneath the patients’ feet and at the forefinger during deep brain stimulation.

Pereira et al. [27], [28] proposed the identification of PD using visual features extracted from a handwriting exam comprised by tasks supposed to be nontrivial for PD patients. The patient is asked to perform some drawings (spirals and meanders) over a guideline that are compared against templates, whose difference between them (tremor) provides information to feed machine learning techniques to classify an individual as either healthy or PD patient. In a later work, Pereira et al. [29], [30] drove their approach to a deep learning application focusing on the signals captured during the same handwriting exams from [27], [28] using a smart pen.1 The signals were mapped into the image domain with different resolutions using visual rhythms and used as input to a Convolutional Neural Network (CNN). Despite the possible loss of information caused by the mapping of the signals, the approach was able to outperform those proposed in [27], [28].

Although the recent advances achieved by computer-aided detection and diagnosis systems, there is still room to improve research in automatic PD identification. Moreover, any improvement is relevant to early diagnosis. As shown in the work of Pereira et al. [29], signals turned to be a promising data source for the PD identification since more subtle variations can be detected. In this work, we propose an extension of the work performed by Pereira et al. [29] by applying the recurrence plot technique (RP) to map signals into the image domain to further feed a CNN. The RP is a technique proposed by Eckmann et al. [31] that enables to visualize repeated events of higher dimension through projections onto a two or three-dimensional representation. Moreover, CNNs are capable of learning very discriminative features, and their application has provided many promising results as well [32].

In short, the contributions of this work are: (i) to the best of our knowledge, this is the first application of recurrence plots for modeling signals in the context of PD identification, (ii) the investigation of the proposed approach performance under different configurations, and (iii) the accuracy improvement provided by the usage of RP with respect to our previous work [29].

The remainder of this work is organized as follows. The proposed approach, as well as material and methods used as evaluation tools, are described in Section 2. The experiments are detailed in Section 3, and Section 4 states conclusions and future works.

Section snippets

Proposed approach and methodology

This section describes the material and methods employed to evaluate the proposed approach for automatic PD patients identification. At a glance, signals were recorded when patients and healthy control (HC) individuals perform some specific drawings. The following step models the signals into images via the recurrence plot method whose output highlights the patterns of each class. Further, a machine learning algorithm performs the classification based on the patterns learned from the output

Experimental results

The proposed approach is evaluated using the Meander and Spiral datasets under different configurations concerning image resolution and training set size. The experimental results are organized by neural architecture and pair [image resolution, training set size] for both Meander and Spiral datasets. The accuracy rates were computed using the standard formulation, i.e., (1#misclassificationstotal#samples)100. The best results of each pair are highlighted in bold according to the Wilcoxon

Conclusions and future works

This work introduced the recurrence plot for modeling signals to automatic Parkinson’s disease identification. The signals are referred as patient’s handwriting dynamics recorded during an exam, which are further used as input to a CNN that identifies whether an individual has or not PD. The main contribution of this work is the improvement of the precision of PD identification using deep learning-based approaches.

Experimental results showed satisfactory and significant improvement of the

Acknowledgments

The authors would like to thank CNPq, Brazil (grants #304315/2017-6, #470501/2013-8, #301928/2014-2, #307066/2017-7, and #306166/2014-3), and FAPESP, Brazil (grants #2013/07375-0, #2014/12236-1, and #2016/19403-6).

Luis Claudio Sugi Afonso is Ph.D. student in the Federal University of São Carlos (UFSCar), and he received his B.Sc. in Information Systems (2010) and the M.Sc in Computer Science (2013) from the São Paulo State University (UNESP). He has experience in Image Processing and Analysis, Artificial Intelligence, and Pattern Recognition.

References (36)

  • PereiraC.R. et al.

    A new computer vision-based approach to aid the diagnosis of parkinson’s disease

    Comput. Methods Programs Biomed.

    (2016)
  • PereiraC.R. et al.

    Handwritten dynamics assessment through convolutional neural networks: An application to parkinson’s disease identification

    Artif. Intell. Med.

    (2018)
  • WangH. et al.

    Breast mass classification via deeply integrating the contextual information from multi-view data

    Pattern Recognit.

    (2018)
  • FariaF.A. et al.

    Fusion of time series representations for plant recognition in phenology studies

    Pattern Recognit. Lett.

    (2016)
  • N.P. Fundation, The parkinson’s foundation’s moving day walks fund local parkinson’s programs in 2017....
  • GelbD. et al.

    Diagnostic criteria for parkinson disease

    Arch. Neurol.

    (1999)
  • ZhaoY.J. et al.

    Factors affecting health-related quality of life amongst asian patients with parkinson’s disease

    Eur. J. Neurol.

    (2008)
  • GoodwillA.M. et al.

    Using non-invasive transcranial stimulation to improve motor and cognitive function in parkinson’s disease: a systematic review and meta-analysis

    Sci. Rep.

    (2017)
  • Cited by (88)

    View all citing articles on Scopus

    Luis Claudio Sugi Afonso is Ph.D. student in the Federal University of São Carlos (UFSCar), and he received his B.Sc. in Information Systems (2010) and the M.Sc in Computer Science (2013) from the São Paulo State University (UNESP). He has experience in Image Processing and Analysis, Artificial Intelligence, and Pattern Recognition.

    Gustavo Henrique de Rosa received his B.Sc. in Computer Science (UNESP, 2016) and he is M.Sc student in Computer Science at UNESP. He has developed research on Image Processing, heuristic optimization techniques, Machine Learning and Pattern Recognition.

    Clayton Reginaldo Pereira is graduated in B.Sc. (2008) from the College Orígenes Lessa, M.Sc in Computer Science (2012) from Universidade Estadual Paulista Julio de Mesquita Filho (UNESP), Ph.D. student (2017) at Universidade Federal de São Carlos (UFSCar). Currently is PostDoc student at UNESP. He has developed research on Image Processing, signal analysis, heuristic optimization techniques, Machine Learning and Pattern Recognition.

    Silke Anna Theresa Weber graduated in Medicine from the São Paulo State University in 1991. She received her MSc and Ph.D. at the São Paulo State University in 2002 and 2006, respectively. She has expertise in Medical Technology, Otolaryngology and is currently Professor in the Medical School of the São Paulo State University, Botucatu.

    Christian Hook received his Ph.D. in the Aachen University. He was a former Professor at Fakultät Informatik und Mathematik at Ostbayerische Technische Hochschule. His main interests are signal processing and machine learning.

    Victor Hugo C. de Albuquerque has a Ph.D. in Mechanical Engineering with emphasis on Materials from the Federal University of Paraíba (UFPB, 2010), an MSc in Teleinformatics Engineering from the Federal University of Ceará (UFC, 2007), and he graduated in Mechatronics Technology at the Federal Center of Technological Education of Ceará (CEFETCE, 2006). He is currently Assistant VI Professor of the Graduate Program in Applied Informatics, and coordinator of the Laboratory of Bioinformatics at the University of Fortaleza (UNIFOR). He has experience in Computer Systems, mainly in the research fields of: Applied Computing, Intelligent Systems, Visualization and Interaction, with specific interest in Pattern Recognition, Artificial Intelligence, Image Processing and Analysis, as well as Automation with respect to biological signal/image processing, image segmentation, biomedical circuits and human/brain–machine interaction, including Augmented and Virtual Reality Simulation Modeling for animals and humans. Additionally, he has research at the microstructural characterization field through the combination of non-destructive techniques with signal/image processing and analysis and pattern recognition. Prof. Victor is the leader of the Computational Methods in Bioinformatics Research Group. He is an editorial board member of the IEEE Access, Computational Intelligence and Neuroscience, Journal of Nanomedicine and Nanotechnology Research, and Journal of Mechatronics Engineering, and he has been Lead Guest Editor of several high-reputed journals, and TPC member of many international conferences. He has authored or coauthored over 160 papers in refereed international journals, conferences, four book chapters, and four patents.

    João P. Papa received his B.Sc. in Information Systems from the São Paulo State University, SP, Brazil. In 2005, he received his M.Sc. in Computer Science from the Federal University of São Carlos, SP, Brazil. In 2008, he received his Ph.D. in Computer Science from the University of Campinas, SP, Brazil. During 2008–2009, he had worked as a post-doctorate researcher at the same institute, and during 2014–2015 he worked as a visiting scholar at Harvard University. He has been a Professor at the Computer Science Department, São Paulo, State University, since 2009, and his research interests include machine learning, pattern recognition, and image processing. He is also the recipient of the Alexander von Humboldt fellowship.

    View full text