Objective Automatic Assessment of Rehabilitative Speech Treatment in Parkinson's Disease

From Prodromal Stages to Clinical Trials: The Promise of Digital Speech Biomarkers in Parkinson's Disease

Speech impairment is a common and disabling symptom in Parkinson's disease (PD), affecting communication and quality of life. Advances in digital speech processing and artificial intelligence have revolutionized objective speech analysis. Given the complex nature of speech impairment, acoustic speech analysis offers unique biomarkers for neuroprotective treatments from the prodromal stages of PD. Digital speech biomarkers can monitor levodopa-induced motor complications, detect the effects of deep brain stimulation, and provide feedback for behavioral speech therapy. This review updates the mechanisms underlying speech impairment, the impact of speech phenotypes, and the effects of interventions on speech. We evaluate the strengths, potential weaknesses, and suitability of promising digital speech biomarkers in PD for capturing disease progression and treatment efficacy. Additionally, we explore the translational potential of PD speech biomarkers to other neuropsychiatric diseases, offering insights into motion, cognition, and emotion. Finally, we highlight knowledge gaps and suggest directions for future research to enhance the use of quantitative speech measures in disease-modifying clinical trials. The findings demonstrate that one year is sufficient to detect disease progression in early PD through speech biomarkers. Voice quality, pitch, loudness, and articulation measures appear to capture the efficacy of treatment interventions most effectively. Certain speech features, such as loudness and articulation rate, behave oppositely in different neurological diseases, offering valuable insights for differential diagnosis. In conclusion, this review highlights speech as a biomarker in tracking disease progression, especially in the prodromal stages of PD, and calls for further longitudinal studies to establish its efficacy across diverse populations.

Towards interpretable speech biomarkers: exploring MFCCs

While speech biomarkers of disease have attracted increased interest in recent years, a challenge is that features derived from signal processing or machine learning approaches may lack clinical interpretability. As an example, Mel frequency cepstral coefficients (MFCCs) have been identified in several studies as a useful marker of disease, but are regarded as uninterpretable. Here we explore correlations between MFCC coefficients and more interpretable speech biomarkers. In particular we quantify the MFCC2 endpoint, which can be interpreted as a weighted ratio of low- to high-frequency energy, a concept which has been previously linked to disease-induced voice changes. By exploring MFCC2 in several datasets, we show how its sensitivity to disease can be increased by adjusting computation parameters.

Applied Machine Learning Techniques to Diagnose Voice-Affecting Conditions and Disorders: Systematic Literature Review

BACKGROUND

Normal voice production depends on the synchronized cooperation of multiple physiological systems, which makes the voice sensitive to changes. Any systematic, neurological, and aerodigestive distortion is prone to affect voice production through reduced cognitive, pulmonary, and muscular functionality. This sensitivity inspired using voice as a biomarker to examine disorders that affect the voice. Technological improvements and emerging machine learning (ML) technologies have enabled possibilities of extracting digital vocal features from the voice for automated diagnosis and monitoring systems.

OBJECTIVE

This study aims to summarize a comprehensive view of research on voice-affecting disorders that uses ML techniques for diagnosis and monitoring through voice samples where systematic conditions, nonlaryngeal aerodigestive disorders, and neurological disorders are specifically of interest.

METHODS

This systematic literature review (SLR) investigated the state of the art of voice-based diagnostic and monitoring systems with ML technologies, targeting voice-affecting disorders without direct relation to the voice box from the point of view of applied health technology. Through a comprehensive search string, studies published from 2012 to 2022 from the databases Scopus, PubMed, and Web of Science were scanned and collected for assessment. To minimize bias, retrieval of the relevant references in other studies in the field was ensured, and 2 authors assessed the collected studies. Low-quality studies were removed through a quality assessment and relevant data were extracted through summary tables for analysis. The articles were checked for similarities between author groups to prevent cumulative redundancy bias during the screening process, where only 1 article was included from the same author group.

RESULTS

In the analysis of the 145 included studies, support vector machines were the most utilized ML technique (51/145, 35.2%), with the most studied disease being Parkinson disease (PD; reported in 87/145, 60%, studies). After 2017, 16 additional voice-affecting disorders were examined, in contrast to the 3 investigated previously. Furthermore, an upsurge in the use of artificial neural network-based architectures was observed after 2017. Almost half of the included studies were published in last 2 years (2021 and 2022). A broad interest from many countries was observed. Notably, nearly one-half (n=75) of the studies relied on 10 distinct data sets, and 11/145 (7.6%) used demographic data as an input for ML models.

CONCLUSIONS

This SLR revealed considerable interest across multiple countries in using ML techniques for diagnosing and monitoring voice-affecting disorders, with PD being the most studied disorder. However, the review identified several gaps, including limited and unbalanced data set usage in studies, and a focus on diagnostic test rather than disorder-specific monitoring. Despite the limitations of being constrained by only peer-reviewed publications written in English, the SLR provides valuable insights into the current state of research on ML-based voice-affecting disorder diagnosis and monitoring and highlighting areas to address in future research.

New Insights into Stroke from Continuous Passively Collected Temperature and Sleep Data Using Wrist-Worn Wearables

Actigraphy may provide new insights into clinical outcomes and symptom management of patients through passive, continuous data collection. We used the GENEActiv smartwatch to passively collect actigraphy, wrist temperature, and ambient light data from 27 participants after stroke or probable brain transient ischemic attack (TIA) over 42 periods of device wear. We computed 323 features using established algorithms and proposed 25 novel features to characterize sleep and temperature. We investigated statistical associations between the extracted features and clinical outcomes evaluated using clinically validated questionnaires to gain insight into post-stroke recovery. We subsequently fitted logistic regression models to replicate clinical diagnosis (stroke or TIA) and disability due to stroke. The model generalization performance was assessed using a leave-one-subject-out cross validation method with the selected feature subsets, reporting the area under the curve (AUC). We found that several novel features were strongly correlated (|r|>0.3) with stroke symptoms and mental health measures. Using selected novel features, we obtained an AUC of 0.766 to estimate diagnosis and an AUC of 0.749 to estimate whether disability due to stroke was present. Collectively, these findings suggest that features extracted from the temperature smartwatch sensor may reveal additional clinically useful information over and above existing actigraphy-based features.

Early detection of Parkinson’s disease from multiple signal speech: Based on Mandarin language dataset

Parkinson’s disease (PD) is a neurodegenerative disorder that negatively affects millions of people. Early detection is of vital importance. As recent researches showed dysarthria level provides good indicators to the computer-assisted diagnosis and remote monitoring of patients at the early stages. It is the goal of this study to develop an automatic detection method based on newest collected Chinese dataset. Unlike English, no agreement was reached on the main features indicating language disorders due to vocal organ dysfunction. Thus, one of our approaches is to classify the speech phonation and articulation with a machine learning-based feature selection model. Based on a relatively big sample, three feature selection algorithms (LASSO, mRMR, Relief-F) were tested to select the vocal features extracted from speech signals collected in a controlled setting, followed by four classifiers (Naïve Bayes, K-Nearest Neighbor, Logistic Regression and Stochastic Gradient Descent) to detect the disorder. The proposed approach shows an accuracy of 75.76%, sensitivity of 82.44%, specificity of 73.15% and precision of 76.57%, indicating the feasibility and promising future for an automatic and unobtrusive detection on Chinese PD. The comparison among the three selection algorithms reveals that LASSO selector has the best performance regardless types of vocal features. The best detection accuracy is obtained by SGD classifier, while the best resulting sensitivity is obtained by LR classifier. More interestingly, articulation features are more representative and indicative than phonation features among all the selection and classifying algorithms. The most prominent articulation features are F1, F2, DDF1, DDF2, BBE and MFCC.

A user-guided Bayesian framework for ensemble feature selection in life science applications (UBayFS)

Feature selection reduces the complexity of high-dimensional datasets and helps to gain insights into systematic variation in the data. These aspects are essential in domains that rely on model interpretability, such as life sciences. We propose a (U)ser-Guided (Bay)esian Framework for (F)eature (S)election, UBayFS, an ensemble feature selection technique embedded in a Bayesian statistical framework. Our generic approach considers two sources of information: data and domain knowledge. From data, we build an ensemble of feature selectors, described by a multinomial likelihood model. Using domain knowledge, the user guides UBayFS by weighting features and penalizing feature blocks or combinations, implemented via a Dirichlet-type prior distribution. Hence, the framework combines three main aspects: ensemble feature selection, expert knowledge, and side constraints. Our experiments demonstrate that UBayFS (a) allows for a balanced trade-off between user knowledge and data observations and (b) achieves accurate and robust results.

Classification of Dysphonic Voices in Parkinson's Disease with Semi-Supervised Competitive Learning Algorithm

This article proposes a novel semi-supervised competitive learning (SSCL) algorithm for vocal pattern classifications in Parkinson’s disease (PD). The acoustic parameters of voice records were grouped into the families of jitter, shimmer, harmonic-to-noise, frequency, and nonlinear measures, respectively. The linear correlations were computed within each acoustic parameter family. According to the correlation matrix results, the jitter, shimmer, and harmonic-to-noise parameters presented as highly correlated in terms of Pearson’s correlation coefficients. Then, the principal component analysis (PCA) technique was implemented to eliminate the redundant dimensions of the acoustic parameters for each family. The Mann−Whitney−Wilcoxon hypothesis test was used to evaluate the significant difference of the PCA-projected features between the healthy subjects and PD patients. Eight dominant PCA-projected features were selected based on the eigenvalue threshold criterion and the statistical significance level (p < 0.05) of the hypothesis test. The SSCL algorithm proposed in this paper included the procedures of the competitive prototype seed selection, K-means optimization, and the nearest neighbor classifications. The pattern classification experimental results showed that the proposed SSCL method can provide the excellent diagnostic performances in terms of accuracy (0.838), recall (0.825), specificity (0.85), precision (0.846), F-score (0.835), Matthews correlation coefficient (0.675), area under the receiver operating characteristic curve (0.939), and Kappa coefficient (0.675), which were consistently better than those results of conventional KNN or SVM classifiers.

Neural network input feature selection using structured l2 − norm penalization

Artificial neural networks are referred to as universal approximators due to their inherent ability to reconstruct complex linear and nonlinear output maps conceived as input-output relationships from data sets. This can be done by reducing large networks via regularization in order to establish compact models containing fewer parameters aimed at describing vital dependencies in data sets. In situations where the data sets contain non-informative input features, devising a continuous, optimal input feature selection technique can lead to improved prediction or classification. We propose a continuous input selection technique through a dimensional reduction mechanism using a ‘structured’ l2 − norm regularization. The implementation is done by identifying the most informative feature subsets from a given data set via an adaptive training mechanism. The adaptation involves introducing a novel, modified gradient approach during training to deal with the non-differentiability associated with the gradient of the structured norm penalty. When the method is applied to process data sets, results indicate that the most informative inputs of artificial neural networks can be selected using a structured l2 − norm penalization.

Multi-scale affinities with missing data: Estimation and applications

Many machine learning algorithms depend on weights that quantify row and column similarities of a data matrix. The choice of weights can dramatically impact the effectiveness of the algorithm. Nonetheless, the problem of choosing weights has arguably not been given enough study. When a data matrix is completely observed, Gaussian kernel affinities can be used to quantify the local similarity between pairs of rows and pairs of columns. Computing weights in the presence of missing data, however, becomes challenging. In this paper, we propose a new method to construct row and column affinities even when data are missing by building off a co-clustering technique. This method takes advantage of solving the optimization problem for multiple pairs of cost parameters and filling in the missing values with increasingly smooth estimates. It exploits the coupled similarity structure among both the rows and columns of a data matrix. We show these affinities can be used to perform tasks such as data imputation, clustering, and matrix completion on graphs.

Relevance, redundancy, and complementarity trade-off (RRCT): A principled, generic, robust feature-selection tool

We present a new heuristic feature-selection (FS) algorithm that integrates in a principled algorithmic framework the three key FS components: relevance, redundancy, and complementarity. Thus, we call it relevance, redundancy, and complementarity trade-off (RRCT). The association strength between each feature and the response and between feature pairs is quantified via an information theoretic transformation of rank correlation coefficients, and the feature complementarity is quantified using partial correlation coefficients. We empirically benchmark the performance of RRCT against 19 FS algorithms across four synthetic and eight real-world datasets in indicative challenging settings evaluating the following: (1) matching the true feature set and (2) out-of-sample performance in binary and multi-class classification problems when presenting selected features into a random forest. RRCT is very competitive in both tasks, and we tentatively make suggestions on the generalizability and application of the best-performing FS algorithms across settings where they may operate effectively.

Scalable and accurate variational Bayes for high-dimensional binary regression models

Modern methods for Bayesian regression beyond the Gaussian response setting are often computationally impractical or inaccurate in high dimensions. In fact, as discussed in recent literature, bypassing such a trade-off is still an open problem even in routine binary regression models, and there is limited theory on the quality of variational approximations in high-dimensional settings. To address this gap, we study the approximation accuracy of routinely-used mean-field variational Bayes solutions in high-dimensional probit regression with Gaussian priors, obtaining novel and practically relevant results on the pathological behaviour of such strategies in uncertainty quantification, point estimation and prediction. Motivated by these results, we further develop a new partially-factorized variational approximation for the posterior distribution of the probit coefficients which leverages a representation with global and local variables but, unlike for classical mean-field assumptions, it avoids a fully factorized approximation, and instead assumes a factorization only for the local variables. We prove that the resulting approximation belongs to a tractable class of unified skew-normal densities that crucially incorporates skewness and, unlike for state-of-the-art mean-field solutions, converges to the exact posterior density as p →∞. To solve the variational optimization problem, we derive a tractable coordinate ascent variational inference algorithm that easily scales to p in the tens of thousands, and provably requires a number of iterations converging to 1 as p →∞. Such findings are also illustrated in extensive empirical studies where our novel solution is shown to improve the approximation accuracy of mean-field variational Bayes for any n and p, with the magnitude of these gains being remarkable in those high-dimensional p &gt; n settings where state-of-the-art methods are computationally impractical.

Things to Consider When Automatically Detecting Parkinson’s Disease Using the Phonation of Sustained Vowels: Analysis of Methodological Issues

Diagnosing Parkinson’s Disease (PD) necessitates monitoring symptom progression. Unfortunately, diagnostic confirmation often occurs years after disease onset. A more sensitive and objective approach is paramount to the expedient diagnosis and treatment of persons with PD (PwPDs). Recent studies have shown that we can train accurate models to detect signs of PD from audio recordings of confirmed PwPDs. However, disparities exist between studies and may be caused, in part, by differences in employed corpora or methodologies. Our hypothesis is that unaccounted covariates in methodology, experimental design, and data preparation resulted in overly optimistic results in studies of PD automatic detection employing sustained vowels. These issues include record-wise fold creation rather than subject-wise; an imbalance of age between the PwPD and control classes; using too small of a corpus compared to the sizes of feature vectors; performing cross-validation without including development data; and the absence of cross-corpora testing to confirm results. In this paper, we evaluate the influence of these methodological issues in the automatic detection of PD employing sustained vowels. We perform several experiments isolating each issue to measure its influence employing three different corpora. Moreover, we analyze if the perceived dysphonia of the speakers could be causing differences in results between the corpora. Results suggest that each independent methodological issue analyzed has an effect on classification accuracy. Consequently, we recommend a list of methodological steps to be considered in future experiments to avoid overoptimistic or misleading results.

Hierarchical Boosting Dual-Stage Feature Reduction Ensemble Model for Parkinson's Disease Speech Data

As a neurodegenerative disease, Parkinson's disease (PD) is hard to identify at the early stage, while using speech data to build a machine learning diagnosis model has proved effective in its early diagnosis. However, speech data show high degrees of redundancy, repetition, and unnecessary noise, which influence the accuracy of diagnosis results. Although feature reduction (FR) could alleviate this issue, the traditional FR is one-sided (traditional feature extraction could construct high-quality features without feature preference, while traditional feature selection could achieve feature preference but could not construct high-quality features). To address this issue, the Hierarchical Boosting Dual-Stage Feature Reduction Ensemble Model (HBD-SFREM) is proposed in this paper. The major contributions of HBD-SFREM are as follows: (1) The instance space of the deep hierarchy is built by an iterative deep extraction mechanism. (2) The manifold features extraction method embeds the nearest neighbor feature preference method to form the dual-stage feature reduction pair. (3) The dual-stage feature reduction pair is iteratively performed by the AdaBoost mechanism to obtain instances features with higher quality, thus achieving a substantial improvement in model recognition accuracy. (4) The deep hierarchy instance space is integrated into the original instance space to improve the generalization of the algorithm. Three PD speech datasets and a self-collected dataset are used to test HBD-SFREM in this paper. Compared with other FR algorithms and deep learning algorithms, the accuracy of HBD-SFREM in PD speech recognition is improved significantly and would not be affected by a small sample dataset. Thus, HBD-SFREM could give a reference for other related studies.

Assessing Parkinson's Disease at Scale Using Telephone-Recorded Speech: Insights from the Parkinson's Voice Initiative

Numerous studies have reported on the high accuracy of using voice tasks for the remote detection and monitoring of Parkinson's Disease (PD). Most of these studies, however, report findings on a small number of voice recordings, often collected under acoustically controlled conditions, and therefore cannot scale at large without specialized equipment. In this study, we aimed to evaluate the potential of using voice as a population-based PD screening tool in resource-constrained settings. Using the standard telephone network, we processed 11,942 sustained vowel /a/ phonations from a US-English cohort comprising 1078 PD and 5453 control participants. We characterized each phonation using 304 dysphonia measures to quantify a range of vocal impairments. Given that this is a highly unbalanced problem, we used the following strategy: we selected a balanced subset (n = 3000 samples) for training and testing using 10-fold cross-validation (CV), and the remaining (unbalanced held-out dataset, n = 8942) samples for further model validation. Using robust feature selection methods we selected 27 dysphonia measures to present into a radial-basis-function support vector machine and demonstrated differentiation of PD participants from controls with 67.43% sensitivity and 67.25% specificity. These findings could help pave the way forward toward the development of an inexpensive, remote, and reliable diagnostic support tool for PD using voice as a digital biomarker.

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

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.

Remote Assessment of Parkinson's Disease Symptom Severity Using the Simulated Cellular Mobile Telephone Network

Telemonitoring of Parkinson's Disease (PD) has attracted considerable research interest because of its potential to make a lasting, positive impact on the life of patients and their carers. Purpose-built devices have been developed that record various signals which can be associated with average PD symptom severity, as quantified on standard clinical metrics such as the Unified Parkinson's Disease Rating Scale (UPDRS). Speech signals are particularly promising in this regard, because they can be easily recorded without the use of expensive, dedicated hardware. Previous studies have demonstrated replication of UPDRS to within less than 2 points of a clinical raters' assessment of symptom severity, using high-quality speech signals collected using dedicated telemonitoring hardware. Here, we investigate the potential of using the standard voice-over-GSM (2G) or UMTS (3G) cellular mobile telephone networks for PD telemonitoring, networks that, together, have greater than 5 billion subscribers worldwide. We test the robustness of this approach using a simulated noisy mobile communication network over which speech signals are transmitted, and approximately 6000 recordings from 42 PD subjects. We show that UPDRS can be estimated to within less than 3.5 points difference from the clinical raters' assessment, which is clinically useful given that the inter-rater variability for UPDRS can be as high as 4-5 UPDRS points. This provides compelling evidence that the existing voice telephone network has potential towards facilitating inexpensive, mass-scale PD symptom telemonitoring applications.

A Mobile Application for Smart Computer-Aided Self-Administered Testing of Cognition, Speech, and Motor Impairment

We present a model for digital neural impairment screening and self-assessment, which can evaluate cognitive and motor deficits for patients with symptoms of central nervous system (CNS) disorders, such as mild cognitive impairment (MCI), Parkinson's disease (PD), Huntington's disease (HD), or dementia. The data was collected with an Android mobile application that can track cognitive, hand tremor, energy expenditure, and speech features of subjects. We extracted 238 features as the model inputs using 16 tasks, 12 of them were based on a self-administered cognitive testing (SAGE) methodology and others used finger tapping and voice features acquired from the sensors of a smart mobile device (smartphone or tablet). Fifteen subjects were involved in the investigation: 7 patients with neurological disorders (1 with Parkinson's disease, 3 with Huntington's disease, 1 with early dementia, 1 with cerebral palsy, 1 post-stroke) and 8 healthy subjects. The finger tapping, SAGE, energy expenditure, and speech analysis features were used for neural impairment evaluations. The best results were achieved using a fusion of 13 classifiers for combined finger tapping and SAGE features (96.12% accuracy), and using bidirectional long short-term memory (BiLSTM) (94.29% accuracy) for speech analysis features.

Bio-inspired dimensionality reduction for Parkinson's disease (PD) classification

Given the demand for developing the efficient Machine Learning (ML) classification models for healthcare data, and the potentiality of Bio-Inspired Optimization (BIO) algorithms to tackle the problem of high dimensional data, we investigate the range of ML classification models trained with the optimal subset of features of PD data set for efficient PD classification. We used two BIO algorithms, Genetic Algorithm (GA) and Binary Particle Swarm Optimization (BPSO), to determine the optimal subset of features of PD data set. The data set chosen for investigation comprises 756 observations (rows or records) taken over 755 attributes (columns or dimensions or features) from 252 PD patients. We employed MaxAbsolute feature scaling method to normalize the data and one hold cross-validation method to avoid biased results. Accordingly, the data is split in to training and testing set in the ratio of 70% and 30%. Subsequently, we employed GA and BPSO algorithms separately on 11 ML classifiers (Logistic Regression (LR), linear Support Vector Machine (lSVM), radial basis function Support Vector Machine (rSVM), Gaussian Naïve Bayes (GNB), Gaussian Process Classifier (GPC), k-Nearest Neighbor (kNN), Decision Tree (DT), Random Forest (RF), Multilayer Perceptron (MLP), Ada Boost (AB) and Quadratic Discriminant Analysis (QDA)), to determine the optimal subset of features (reduction of dimensionality) contributing to the highest classification accuracy. Among all the bio-inspired ML classifiers employed: GA-inspired MLP produced the maximum dimensionality reduction of 52.32% by selecting only 359 features and delivering 85.1% of the classification accuracy; GA-inspired AB delivered the maximum classification accuracy of 90.7% producing the dimensionality reduction of 41.43% by selecting only 441 features; And, BPSO-inspired GNB produced the maximum dimensionality reduction of 47.14% by selecting 396 features and delivering the classification accuracy of 79.3%; BPSOMLP delivered the maximum classification accuracy of 89% and produced 46.48% of the dimensionality reduction by selecting only 403 features.

Binary Credal Classification Under Sparsity Constraints

Binary classification is a well known problem in statistics. Besides classical methods, several techniques such as the naive credal classifier (for categorical data) and imprecise logistic regression (for continuous data) have been proposed to handle sparse data. However, a convincing approach to the classification problem in high dimensional problems (i.e., when the number of attributes is larger than the number of observations) is yet to be explored in the context of imprecise probability. In this article, we propose a sensitivity analysis based on penalised logistic regression scheme that works as binary classifier for high dimensional cases. We use an approach based on a set of likelihood functions (i.e. an imprecise likelihood, if you like), that assigns a set of weights to the attributes, to ensure a robust selection of the important attributes, whilst training the model at the same time, all in one fell swoop. We do a sensitivity analysis on the weights of the penalty term resulting in a set of sparse constraints which helps to identify imprecision in the dataset.

Complexity Measures of Voice Recordings as a Discriminative Tool for Parkinson's Disease

In this paper, we have investigated the differences in the voices of Parkinson's disease (PD) and age-matched control (CO) subjects when uttering three phonemes using two complexity measures: fractal dimension (FD) and normalised mutual information (NMI). Three sustained phonetic voice recordings, /a/, /u/ and /m/, from 22 CO (mean age = 66.91) and 24 PD (mean age = 71.83) participants were analysed. FD was first computed for PD and CO voice recordings, followed by the computation of NMI between the test groups: PD-CO, PD-PD and CO-CO. Four features reported in the literature-normalised pitch period entropy (Norm. PPE), glottal-to-noise excitation ratio (GNE), detrended fluctuation analysis (DFA) and glottal closing quotient (ClQ)-were also computed for comparison with the proposed complexity measures. The statistical significance of the features was tested using a one-way ANOVA test. Support vector machine (SVM) with a linear kernel was used to classify the test groups, using a leave-one-out validation method. The results showed that PD voice recordings had lower FD compared to CO (p < 0.008). It was also observed that the average NMI between CO voice recordings was significantly lower compared with the CO-PD and PD-PD groups (p < 0.036) for the three phonetic sounds. The average NMI and FD demonstrated higher accuracy (>80%) in differentiating the test groups compared with other speech feature-based classifications. This study has demonstrated that the voices of PD patients has reduced FD, and NMI between voice recordings of PD-CO and PD-PD is higher compared with CO-CO. This suggests that the use of NMI obtained from the sample voice, when paired with known groups of CO and PD, can be used to identify PD voices. These findings could have applications for population screening.

Quantifying ultrasonic mouse vocalizations using acoustic analysis in a supervised statistical machine learning framework

Examination of rodent vocalizations in experimental conditions can yield valuable insights into how disease manifests and progresses over time. It can also be used as an index of social interest, motivation, emotional development or motor function depending on the animal model under investigation. Most mouse communication is produced in ultrasonic frequencies beyond human hearing. These ultrasonic vocalizations (USV) are typically described and evaluated using expert defined classification of the spectrographic appearance or simplistic acoustic metrics resulting in nine call types. In this study, we aimed to replicate the standard expert-defined call types of communicative vocal behavior in mice by using acoustic analysis to characterize USVs and a principled supervised learning setup. We used four feature selection algorithms to select parsimonious subsets with maximum predictive accuracy, which are then presented into support vector machines (SVM) and random forests (RF). We assessed the resulting models using 10-fold cross-validation with 100 repetitions for statistical confidence and found that a parsimonious subset of 8 acoustic measures presented to RF led to 85% correct out-of-sample classification, replicating the experts’ labels. Acoustic measures can be used by labs to describe USVs and compare data between groups, and provide insight into vocal-behavioral patterns of mice by automating the process on matching the experts’ call types.

Screening and Diagnosis of Chronic Pharyngitis Based on Deep Learning

Chronic pharyngitis is a common disease, which has a long duration and a wide range of onset. It is easy to misdiagnose by mistaking it with other diseases, such as chronic tonsillitis, by using common diagnostic methods. In order to reduce costs and avoid misdiagnosis, the search for an affordable and rapid diagnostic method is becoming more and more important for chronic pharyngitis research. Speech disorder is one of the typical symptoms of patients with chronic pharyngitis. This paper introduces a convolutional neural network model for diagnosis based on the typical symptom of speech disorder. First of all, the voice data is converted into a speech spectrogram, which can better output the speech characteristic information and lay a foundation for computer diagnosis and discrimination. Second, we construct a deep convolutional neural network for the diagnosis of chronic pharyngitis through the design of the structure, the design of the network layer, and the description of the function. Finally, we perform a parameter optimization experiment on the convolutional neural network and judge the recognition efficiency of chronic pharyngitis. The results show that the convolutional neural network has a high recognition rate for patients with chronic pharyngitis and has a good diagnostic effect.

Developing a large scale population screening tool for the assessment of Parkinson's disease using telephone-quality voice

Recent studies have demonstrated that analysis of laboratory-quality voice recordings can be used to accurately differentiate people diagnosed with Parkinson's disease (PD) from healthy controls (HCs). These findings could help facilitate the development of remote screening and monitoring tools for PD. In this study, 2759 telephone-quality voice recordings from 1483 PD and 15 321 recordings from 8300 HC participants were analyzed. To account for variations in phonetic backgrounds, data were acquired from seven countries. A statistical framework for analyzing voice was developed, whereby 307 dysphonia measures that quantify different properties of voice impairment, such as breathiness, roughness, monopitch, hoarse voice quality, and exaggerated vocal tremor, were computed. Feature selection algorithms were used to identify robust parsimonious feature subsets, which were used in combination with a random forests (RFs) classifier to accurately distinguish PD from HC. The best tenfold cross-validation performance was obtained using Gram-Schmidt orthogonalization and RF, leading to mean sensitivity of 64.90% (standard deviation, SD, 2.90%) and mean specificity of 67.96% (SD 2.90%). This large scale study is a step forward toward assessing the development of a reliable, cost-effective, and practical clinical decision support tool for screening the population at large for PD using telephone-quality voice.

Towards an automatic evaluation of the dysarthria level of patients with Parkinson's disease

BACKGROUND

Parkinson's disease (PD) is a neurological disorder that produces motor and non-motor impairments. The evaluation of motor symptoms is currently performed following the third section of the Movement Disorder Society - Unified Parkinson's Disease Rating Scale (MDS-UPDRS-III); however, only one item of that scale is related to speech impairments. It is necessary to develop a specific scale such that considers those aspects related to speech impairments of the patients.

AIMS

(i) To introduce and evaluate the suitability of a modified version of the Frenchay Dysarthria Assessment (m-FDA) scale to quantify the dysarthria level of PD patients; (ii) to objectively model dysarthric speech signals considering four speech dimensions; (iii) to develop a methodology, based on speech processing and machine learning methods, to automatically quantify/predict the dysarthria level of patients with PD.

METHODS

The speech recordings are modeled using features extracted from several dimensions of speech: phonation, articulation, prosody, and intelligibility. The dysarthria level is quantified using linear and non-linear regression models. Speaker models based on i-vectors are also explored.

RESULTS AND CONCLUSIONS

The m-FDA scale was introduced to assess the dysarthria level of patients with PD. Articulation features extracted from continuous speech signals to create i-vectors were the most accurate to quantify the dysarthria level, with correlations of up to 0.69 between the predicted m-FDA scores and those assigned by the phoniatricians. When the dysarthria levels were estimated considering dedicated speech exercises such as rapid repetition of syllables (DDKs) and read texts, the correlations were 0.64 and 0.57, respectively. In addition, the combination of several feature sets and speech tasks improved the results, which validates the hypothesis about the contribution of information from different tasks and feature sets when assessing dysarthric speech signals. The speaker models seem to be promising to perform individual modeling for monitoring the dysarthria level of PD patients. The proposed approach may help clinicians to make more accurate and timely decisions about the evaluation and therapy associated to the dysarthria level of patients. The proposed approach is a great step towards unobtrusive/ecological evaluations of patients with dysarthric speech without the need of attending medical appointments.

Smartphone motor testing to distinguish idiopathic REM sleep behavior disorder, controls, and PD

Objective

We sought to identify motor features that would allow the delineation of individuals with sleep study-confirmed idiopathic REM sleep behavior disorder (iRBD) from controls and Parkinson disease (PD) using a customized smartphone application.

Methods

A total of 334 PD, 104 iRBD, and 84 control participants performed 7 tasks to evaluate voice, balance, gait, finger tapping, reaction time, rest tremor, and postural tremor. Smartphone recordings were collected both in clinic and at home under noncontrolled conditions over several days. All participants underwent detailed parallel in-clinic assessments. Using only the smartphone sensor recordings, we sought to (1) discriminate whether the participant had iRBD or PD and (2) identify which of the above 7 motor tasks were most salient in distinguishing groups.

Results

Statistically significant differences based on these 7 tasks were observed between the 3 groups. For the 3 pairwise discriminatory comparisons, (1) controls vs iRBD, (2) controls vs PD, and (3) iRBD vs PD, the mean sensitivity and specificity values ranged from 84.6% to 91.9%. Postural tremor, rest tremor, and voice were the most discriminatory tasks overall, whereas the reaction time was least discriminatory.

Conclusions

Prodromal forms of PD include the sleep disorder iRBD, where subtle motor impairment can be detected using clinician-based rating scales (e.g., Unified Parkinson's Disease Rating Scale), which may lack the sensitivity to detect and track granular change. Consumer grade smartphones can be used to accurately separate not only iRBD from controls but also iRBD from PD participants, providing a growing consensus for the utility of digital biomarkers in early and prodromal PD.