Monitoring Parkinson's disease progression based on recorded speech with missing ordinal responses and replicated covariates

https://doi.org/10.1016/j.compbiomed.2021.104503Get rights and content

Abstract

Monitoring Parkinson's Disease (PD) progression is an important task to improve the life quality of the affected people. This task can be performed by extracting features from voice recordings and applying specifically designed statistical models, leading to systems that improve the ability of monitoring the progression of PD in an objective, remote, non-invasive, fast, and economically sustainable way. An experiment has been conducted with 36 subjects to study the progression of the PD over 4 years by using the Hoehn and Yahr (HY) scale and features extracted from the phonation of the vowel/a/. The collected dataset had many missing data, which should be addressed jointly with the non-decreasing nature of the disease and the within-subject variability due to the use of replicated features. In order to handle these issues, a Hidden Markov model for longitudinal data was designed and implemented by using a data augmentation scheme based on different latent variables. Markov chain Monte Carlo methods were used to generate from the posterior distribution. The proposed approach has been tested on simulated data, providing good accuracy rates in the context of a multiclass problem. It also has been applied to the real data obtained from the conducted experiment, providing imputed and predicted HY stages compatible with the progression of PD. The conducted experiment and the proposed approach contribute to fill a gap in the scientific literature on experiments and methodologies for tracking PD progression based on acoustic features and the HY scale. This would help to derive an expert system that can be integrated into the protocols of neurology units in hospital centers.

Introduction

Parkinson's Disease (PD) is a long-term neurodegenerative disorder that mainly affects the motor system with symptoms including tremor, stiffness, instability, lack of coordination, or difficulty with walking. Non-motor symptoms such as cognitive and behavioral problems are also relevant. Besides, voice production is affected by motor and cognitive problems. These symptoms begin gradually and get worse over the time.

According to the Parkinson's Disease Foundation, it is estimated that PD currently affects 7 to 10 million people worldwide, being the most relevant neurodegenerative disorder after Alzheimer's disease, but with a faster growth. The Global Burden of Disease study projects to reach 13 million people affected by 2040 [1]. This gives an idea of the magnitude of the problem and justifies the great effort in research and medical care to improve the life quality of people suffering from this up-to-now incurable disorder.

The early diagnosis of PD is key to improve the life quality of people who suffer from it. The diagnosis of this disorder is not evident and takes time (between 1 and 3 years). Besides, it requires the intervention of specialized neurologists. Tracking PD progression is also very important. Receiving continuous monitoring of the progression of the disease is especially interesting, since the symptoms fluctuate significantly throughout the day and, in general conditions, the neurologist can only assess the patient's situation at the specific time of day in which the physical consultation is carried out (once a year in many public health systems). The dose of medication and its administration can be customized according to the evolution of the patient's symptoms.

In recent years, Computer-Aided Diagnosis (CAD) systems have been built to aid in the detection and monitoring of many diseases [2], and, in particular, those detectable by voice [3]. Since voice production is affected by PD, CAD systems, based on features extracted from voice recordings, can be used for these tasks. Moro-Velazquez et al. [4] present a recent review of the advances in PD detection and assessment using voice. Many approaches have been derived for PD detection, according to different experiments involving different phonation tasks (/a/sustained phonation [5]; /ka/syllable phonation [6]; and phonations of isolated words, rapid repetition of the syllables, sentences, and read texts [7]), feature extraction procedures (Jitter, Shimmer, MFCCs, HNR, RPDE, DFA, Entropies … [4]), or classification approaches (k-nearest neighbor, random forests, gradient boosting … [8]). However, the number of studies for tracking PD progression is much more limited, as there is an underlying difficulty in conducting long-term studies. Furthermore, regression-based approaches for longitudinal studies, addressing the difficulties related to the particular experimental designs, are more difficult to derive and apply than classification models for cross-sectional studies. Regression analysis is a set of statistical techniques used to evaluate the relationship among variables, i.e., it tries to determine if one or more independent variables can explain the variability produced in a dependent variable (response variable) [9]. Regression models for numerical and ordinal response are especially interesting in this longitudinal context.

PD can be monitored by using the Unified Parkinson's Disease Rating Scale (UPDRS) [10]. UPDRS is a numerical scale used to measure the course of PD. It enables the quantification of the type, number, and severity of extrapyramidal signs. This quantification is partially based on subjective criteria, therefore disagreement among raters in the interpretation of these criteria may happen [11]. UPRDS contains 45 questions for rating, divided into four parts: (I) Mental state (16 points); (II) Daily life activities (52 points); (III) Motor aspects (68 points); and (IV) Complication of treatments (23 points). Some studies have found approximations to UPDRS scores from regression models applied to features extracted from voice recordings both from linear and nonlinear regression techniques. Tsanas et al. [12] used classical least squares regression and regression trees, whereas Hemmerling and Wojcik-Pedziwiatr [13] used multiple linear regression, random forest regression and support vector machine regression for the same purpose. However, most of the approaches carried out in the scientific literature have been developed on the basis of a single multicenter study conducted in the United States with a total of 52 subjects with PD during a short period of 6 months [14]. The voice recordings in this study are not publicly available, but the features, which were extracted for a subgroup of 42 patients, can be downloaded from the UCI Machine Learning Repository.1 Based on this dataset, some approaches have been developed and applied. Besides Tsanas et al. [12], Eskidere et al. [15] used this dataset for model performance assessment with regression models based on support vector machine, least square support vector machines, multilayer perceptron neural network, and general neural network regression methods. Naranjo et al. [16] developed a binary regression model that addressed voice recording replications. Nilashi et al. [17] developed an approach based on ensembles of deep belief network and self-organizing map.

The application of UPDRS scale (or MDS-UPDRS [18]) requires the presence of the patient in a hospital center, as well as extensive physical examinations by qualified medical personnel. Besides, the inter-rater agreement is not as good as desirable [19]. The Hoehn and Yahr (HY) scale is simpler and used very often to assess the level of disability produced by PD [20]. Originally, it contained 5 stages on an ordinal scale, but later was modified by including two more stages to help in describing the intermediate course of the disease [21]. Skodda et al. [22] conducted a longitudinal study to assess the progression and speech impairment in the course of PD based on the UPDRS motor and HY scales, although they simplified the HY scale to only 3 stages. They focused on analyzing the correlation between perceptual speech scores (articulation, fluency, prosody …) and basic linear speech parameters, and they made a comparison between PD and control groups. They did not try to make predictions based on the acoustic features.

There is a lack of longitudinal studies to predict HY stages based on acoustic features. This kind of prediction can be performed based on a regression model for ordinal response. This has motivated conducting an experiment with 36 people suffering from PD over 4 years, with stages ranging from 1 to 4 (no subject reached stage 5). Carrying out this experiment led to some challenges that required the construction of a specific approach. The first challenge was the existence of non-response. The experiment was conducted in the headquarters of the Regional Association for Parkinson's Disease of Extremadura (Spain), and the participating subjects were volunteers, which were invited to participate every year. Some of them were not available during some periods of time (illness, travels …) or even permanently. Therefore, missing data were obtained in the following rates: 11% after one year, 22% after two years, and 39% after three years. In spite of that, a great amount of informative data was collected and it was needed to build appropriate statistical methodology to address it. Besides, PD is a neurodegenerative disorder, so the people that suffer from it get progressively worse over the time. This situation also requires to address the problem as a non-decreasing process. Many diseases have been modelled taking this fact into account such as caries experience [23], aortic aneurysm [24], or even PD [25].

Addressing the issues of non-response and non-decreasing process should be completed with managing the within-subject variability produced by the use of replicated voice recordings. There exists a relevant variability between features extracted from two or more voice recordings of the same subject at a particular time, so using only one utterance per subject may provide different results depending on the voice recording that has been selected. Imperfections in technology and the very biological variability result in values that are similar (but not identical) for voice recordings from a particular subject, rather than for recordings from different individuals. Some authors have ignored this issue and have treated the data as if they were independent when there exists a clear dependent nature (see, e.g., the treatment given by Tsanas et al. [12] or Nilashi et al. [17]). However, this within-subject variability can be properly addressed [25]. The experiment involved in this article considers three replications per subject at each time point.

In this article, a Hidden Markov Model (HMM) has been developed and applied to address the lack of response in the HY ordinal scale for tracking PD progression, non-decreasing process, and within-subject variability produced by the voice recording replications. Due to the difficulty of addressing these issues, latent variables have been introduced in the model, which has allowed to solve the model by using Markov chain Monte Carlo (MCMC) methods. The resulting approach has been applied to the data collected from the previously described long-term experiment, which has been conducted specifically for this task. Therefore, the proposed approach is part of a CAD system that contributes to improve the ability of monitoring the progression of PD in an objective, remote, non-invasive, fast, and economically sustainable way. This is in line with the development and use of tools, technologies and digital solutions for health and care, as an essential pillar of modern medicine. Although the approach has been derived for monitoring the progression of PD, it is applicable to other problems that share these characteristics.

The outline of the rest of the article is as follows. Section 2 presents the data collection through the description of the participants, speech recordings, recording devices, and feature extraction procedures. Section 3 describes the HMM, including the details about dealing with the missing ordinal response, non-decreasing process, and replicated covariates. Also, the posterior distribution is explored. Section 4 presents the results obtained from a simulation-based case and from real data obtained with the conducted experiment. Next, conclusions are shown in Section 5. Finally, imputed and predicted stages are presented in an Appendix.

Section snippets

Participants

A total of 36 subjects having PD have been involved in this study, being 12 women (33.3%) and 24 men (66.7%). The mean (standard deviation) age was 69 (7.93) years at the beginning of the study. All had a definitive diagnosis by their neurologists and were medicated with levodopa. They attended different activities at the headquarters of the Regional Association for Parkinson's Disease of Extremadura in Cáceres and Mérida (Spain) during the years 2016–2019.

The modified HY scale was used to

Approach

In this section, the HMM-based approach is proposed. It is composed of several parts. Firstly, the ordinal regression model is formulated. Then, a data augmentation framework is considered to address the monotone non-decreasing process. Next, it is explained how the replications in the explanatory variables are integrated in the model, and how the missing data are addressed. Finally, the prior distribution is presented and the posterior distribution is derived. Using MCMC methods to solve the

Results

Firstly, a simulation-based case is presented to show the model performance. Then, the approach is applied to the real data coming from the conducted experiment and the results are displayed as well as discussed.

Conclusion

Tracking the progression of PD can be addressed with the use of acoustic features in an objective, remote, non-invasive, fast, and economically sustainable way. Some approaches have been used for the UPDRS scale, but there is a lack of studies and approaches for the HY scale. The conducted experiment provided challenging data that must be addressed with a specific approach that is able to manage missing data, the non-decreasing nature of the disease, and replicated covariates obtained from

Acknowledgments

This research has been supported by Agencia Estatal de Investigación, Spain (Project MTM2017-86875-C3-2-R), Junta de Extremadura, Spain (Projects IB16054, GR18108, and GR18055), and the European Union (European Regional Development Funds). Lizbeth Naranjo has also been partially supported by UNAM-DGAPA-PAPIIT, Mexico (Project IN118720).

All procedures performed in this study involving human participants were in accordance with the ethical standards of the Bioethical Committee of the University

References (41)

  • J.R. Orozco-Arroyave et al.

    Automatic detection of Parkinson's disease in running speech spoken in three different languages

    J. Acoust. Soc. Am.

    (2016)
  • D. Jain et al.

    Machine learning based automatic prediction of Parkinson's disease using speech features

  • S. Chatterjee et al.
  • S. Fahn et al.
    (1987)
  • B. Post et al.

    Unified Parkinson's disease rating scale motor examination: are ratings of nurses, residents in neurology, and movement disorders specialists interchangeable?

    Mov. Disord.

    (2005)
  • A. Tsanas et al.

    Accurate telemonitoring of Parkinson's disease progression by non-invasive speech tests

    IEEE Trans. Biomed. Eng.

    (2010)
  • D. Hemmerling et al.

    Prediction and estimation of Parkinson's disease severity based on voice signal

    J. Voice

    (2020)
  • C.G. Goetz et al.

    Testing objective measures of motor impairment in early Parkinson's disease: feasibility study of an at-home testing device

    Mov. Disord.

    (2009)
  • L. Naranjo et al.

    Addressing voice recording replications for tracking Parkinson's disease progression

    Med. Biol. Eng. Comput.

    (2017)
  • C. Goetz et al.

    Movement disorder society-sponsored revision of the unified Parkinson's disease rating scale (MDS-UPDRS): process, format, and clinimetric testing

    Mov. Disord.

    (2007)
  • Cited by (0)

    View full text