Exploring Feature Selection and Feature Transformation Techniques

to Improve Telephone-based Biomedical Speech Signal Processing

towards Parkinson’s Assessment

Athanasios Tsanas

and Siddharth Arora

Usher Institute, Edinburgh Medical School, University of Edinburgh, U.K.

Department of Mathematics, University of Oxford, U.K.

Keywords: Acoustic Analysis, Parkinson’s Disease, Speech Signal Processing, Sustained Vowels.

Abstract: Clinical decision support tools mining speech signals for Parkinson’s Disease (PD) applications typically rely

on relatively small numbers of participants, having collected data under highly controlled acoustic conditions.

We recently reported on the Parkinson’s Voice Initiative (PVI), a large international project leading to the

collection of 19,000+ sustained vowel phonations (control and PD groups) across seven countries, where

participants were self-selected and provided phonations over the standard telephone network. In this study,

we explored sustained vowels in a balanced subset of the US-speaking cohort in PVI comprising 3000

participants (1500 PD and 1500 controls). The aim was to investigate feature selection and feature

transformation techniques towards improving binary differentiation of controls and PD and obtaining new

insights in a lower dimensional space. We acoustically characterized each sustained vowel /a/ phonation using

307 dysphonia measures which had previously been successfully employed in speech-PD applications. We

explored five different feature selection and two manifold embedding techniques to project data into new

feature spaces which might be more predictive of the binary outcome, and presented those into a Random

Forest. We assessed the performance of the resulting models using internal 10-fold Cross-Validation (CV).

We report classification accuracy of 67% and provide tentative interpretation by comparing the different

feature subsets identified using different methods. Collectively, these findings provide new insights towards

developing parsimonious feature subsets to facilitate the development of a large-scale tool for PD screening

at minimal cost using telephone-based sustained vowels.

1 INTRODUCTION

Parkinson’s Disease (PD) is a progressive

neurodegenerative disorder straining national health

systems globally (Dorsey et al., 2013). Prevalence

rates have been constantly increasing over the last

years: there were approximately 2.5 million People

diagnosed with PD (PwP) in 1990, rising to 6.1

million PwP by 2016 (GBD, 2018). More recently, a

large global burden of disease study highlighted PD

as one of the top five leading causes of death from

neurological disorders in the US (GBD Neurological

Disorders Collaborators, 2021). Cardinal PD

symptoms include tremor, rigidity, bradykinesia, and

postural stability, within the broader remit of motor,

https://orcid.org/0000-0002-0994-8100

https://orcid.org/0000-0001-6499-6941

cognitive, and neuropsychiatric symptoms (Olanow,

Stern, Sethi 2009).

The use of speech signals to assess PD has been

very well described in the research literature (Titze,

2000; Tsanas 2012). It is revealing that 29% of PwP

consider vocal performance degradation as one of

their most debilitating symptoms (Hartelius and

Svensson, 1994). Recent studies have demonstrated

the enormous potential of capitalizing on speech

signals in neurodegenerative applications and PD in

particular. For example, research work has explored:

(1) differentiating PwP from age- and gender-

matched controls with almost 99% accuracy (Tsanas

et al., 2012); (2) accurately replicating the Unified

Parkinson’s Disease Rating Scale (UPDRS) (Tsanas

et al., 2010a; 2010b; 2010c; 2011; 2021), which is the

Tsanas, A. and Arora, S.

Exploring Feature Selection and Feature Transformation Techniques to Improve Telephone-based Biomedical Speech Signal Processing towards Parkinson’s Assessment.

DOI: 10.5220/0011029800003123

In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - Volume 4: BIOSIGNALS, pages 327-334

ISBN: 978-989-758-552-4; ISSN: 2184-4305

327

standard clinical tool to provide an overall PD

symptom assessment; (3) automatically assessing

voice rehabilitation (Tsanas et al., 2014a); (4)

providing early biomarkers in PD with gene

mutations and other PD precursors (Arora et al.,

2018; Arora et al. 2021); (5) clustering PD

participants towards developing more personalized

monitoring and treatment approaches (Tsanas and

Arora, 2020; 2021); and (6) speech articulation

kinematic models to characterize PD dysarthria thus

providing tentative insights into the underlying

physiology (Gomez et al., 2019).

Typically, reseach into speech-PD has focused on

single-site findings and has been limited in terms of

study paticipants. A large multi-site trial, the

Parkinson’s Voice Initiative (PVI) (Arora, Baghai-

Ravary, Tsanas, 2019; Arora and Tsanas, 2021) is the

first of its kind study, inviting people to self-enrol and

donate their voices to facilitate large scale analysis of

PD. Overall, PVI collected more than 19,000

sustained vowel /a/ samples from people across seven

countries. Although the data collected in PVI is

clearly not of the same high quality as data collected

under carefully controlled acoustic conditions, the

large number of samples facilitates new explorations

in different directions.

The aim of this study was to explore different

feature selection and feature transformation

techniques towards facilitating the binary

differentiation of control participants and PD

participants in a subset of the PVI data, thus building

on our previous work with this dataset (Arora,

Baghai-Ravary and Tsanas 2019; Arora and Tsanas,

2021a). The ultimate goal is to develop a clinical

decision support tool to facilitate PD screening at

large at practically no cost.

2 DATA

The PVI study invited people call on a dedicated

region-specific phone number and contribute their

voices to facilitate clinical research into PD. Data

were collected across seven major geographical

locations (Argentina, Brazil, Canada, Mexico, Spain,

USA, and the UK) using servers by Aculab for the

needs of this project. People called a dedicated phone

number that was closest to their geographical location

and were not compensated in any way for

participating in the study. Participants followed aural

instructions in the native language for the region, and

were asked to provide basic demographic information

(age, gender), self-report whether they had received a

clinical PD diagnosis, and record two sustained vowel

/a/ phonations. The instruction was to sustain vowel

/a/ for as long and as steadily as possible, following

standard widely speech collection protocols (Titze,

2000). The speech recordings were sampled at 8 kHz

at 16 bits or resolution. In total, the PVI study

collected more than 19,000 phonations.

In this study we processed data from the single

largest collection site, Boston to ovecome differences

in voices from people coming from different

linguistic backgrounds, even when comparing UK-

English and US-English (Tsanas and Arora, 2021b).

Specifically, we processed data from 1078 PD

participants (age 62.7±12.0, 566 male) and 5453

controls (49.2±15.9, 2976 male). We do not have

detailed information regarding PD-symptom specific

aspects, for example whether participants self-

enrolled when they were “on” or “off” medication, or

clinically validated metrics such as UPDRS. For

further details on PVI including detailed

demographics we refer readers to our previous work

(Arora, Baghai-Ravary, Tsanas, 2019; Tsanas and

Arora, 2019; Arora and Tsanas 2021).

3 METHODS

3.1 Data Pre-Processing

We developed a speech recognition software which

automatically transcribed the participants’ responses

over the phone regarding age, gender, and self-

reported PD assessment. Randomly selected

recordings were aurally inspected for voice quality to

ensure the transcription was correct. Moreover, we

inspected recordings where the automated speech

recognition algorithm had less than 90% confidence

in the transcript output. For further details on

preprocessing and removing faulty phonations we

refer to (Arora, Baghai-Ravary, Tsanas, 2019).

3.2 Acoustic Characterization of

Sustained Vowel /a/ Phonations

We used the Voice Analysis Toolbox (freely

available from https://www.darth-group.com/

software and also from https://github.com/Thanasis

Tsanas/VoiceAnalysisToolbox) to acoustically

characterize each sustained vowel /a/ phonation. The

toolbox computes 307 dysphonia measures, which

have been developed specifically to characterize

sustained vowel /a/ phonations extensively validated

across diverse PD datasets (Tsanas et al., 2010a;

Tsanas et al., 2010b; Tsanas et al., 2011; Tsanas et al.,

SERPICO 2022 - Special Session on Diagnostic, Prognostic, and Phenotyping Models from Mined Administrative Healthcare Data

328

Table 1: Breakdown of the dysphonia measures used in the study.

Family of acoustic measures Brief description

umber o

measures

Jitter variants F0 perturbation 28

Shimmer variants Amplitude perturbation 21

Harmonics to Noise Ratio (HNR) and

Noise to Harmonics Ratio (NHR)

Signal to noise, and noise to signal ratios computed using

standard approaches relying on autocorrelation

Glottis Quotient (GQ)

Vocal fold cycle duration changes, quantified by focusing

on specific glottal opening and glottal closure periods,

quantified using DYPSA (Naylor et al., 2007)

Glottal to Noise Excitation (GNE)

Extent of noise in speech using energy and nonlinear energy

concepts

Vocal Fold Excitation Ratio (VFER)

Extent of noise in speech using energy, nonlinear energy,

and entropy concepts

Empirical Mode Decomposition

Excitation Ratio (EMD-ER)

Signal to noise ratios using EMD-based energy, nonlinear

energy, and entropy

Mel Frequency Cepstral Coefficients

(MFCC)

Amplitude and spectral fluctuations on the Mel scale

quantifying envelope and high frequency aspects

F0 related

Comparisons of F0 against age and gender matched

controls, including probabilistic variabilities

Wavelet-based coefficients

Amplitude, scale, and envelope fluctuations quantified

using wavelet coefficients, and processing with entropy,

Teager-Kaiser Energy, signal energy, and signal to noise

ratios

182

Pitch Period Entropy (PPE)

Variability of F0 expressing inefficiency of F0 stability

over and above the variability exhibited by healthy controls

Detrended Fluctuation Analysis (DFA) Stochastic self-similarity of turbulent noise 1

Recurrence Period Density Entropy

(RPDE)

Uncertainty in estimation of F0 1

Algorithmic expressions for the dysphonia measures summarized above are described in detail in (Tsanas, 2012; Tsanas, 2013). The

MATLAB source code for the computation of the dysphonia measures is freely available from https://www.darth-group.com/software and

also from https://github.com/ThanasisTsanas/VoiceAnalysisToolbox). F0 refers to fundamental frequency estimates, here computed using

SWIPE (Camacho and Harris, 2008).

2012; Tsanas, 2012; Tsanas et al., 2014a; Arora,

Baghai-Ravary, Tsanas, 2019; Tsanas et al., 2021),

and other applications, e.g. processing voice fillers

(Tsanas and Gomez-Vilda, 2013; San Segundo,

Tsanas, Gomez-Vilda, 2017). We have described in

detail previously the background, rationale, and

detailed algorithmic expressions for the computation

of the dysphonia measures (Tsanas, 2012; Tsanas,

2013). A concise summary of the extracted dysphonia

measures is summarized in Table 1 including the

number of dysphonia measures for each algorithmic

family and a brief description.

The fundamental frequency (F0) is a key speech

characteristic, and its estimation is a prerequisite for

the computation of many dysphonia measures, e.g.

for jitter, and Pitch Period Entropy (PPE). There are

many algorithms in the research literature for F0

estimation (Roark, 2006; Tsanas et al., 2014b); in this

study, we used the Sawtooth Waveform Inspired

Pitch Estimator (SWIPE) algorithm (Camacho and

Harris, 2008), which we had previously demonstrated

is the most accurate F0 estimation algorithm for

sustained vowel /a/ phonations (Tsanas et al., 2014b).

Applying the dysphonia measures to each

sustained vowel /a/ phonation gives rise to 307

features which are continuous random variables.

Therefore, we have a 11,942×304 data matrix that we

aim to process further to map onto the binary outcome

(0 was used to denote controls and 1 to denote PwP).

Exploring Feature Selection and Feature Transformation Techniques to Improve Telephone-based Biomedical Speech Signal Processing

towards Parkinson’s Assessment

329

3.3 Dimensionality Reduction

A high dimensional dataset may lead to statistical

learning performance degradation and obfuscates the

understanding of clear patterns in a dataset. This well-

known problem is often referred to as the curse of

dimensionality (Guyon et al. 2006; Hastie, Tibshirani,

Friedman, 2009). Following Occam’s razor, we would

prefer a predictive model which is as simple as

possible, i.e. with a low dimensionality. This approach

is typically referred to as dimensionality reduction, and

can be achieved either by feature transformation

(transforming the features to populate a new, lower

dimensional space), or by feature selection (choosing a

subset of features). Feature selection is often more

suitable in clinical settings to retain the interpretability

of the original features (Guyon et al., 2006; Tsanas,

Little, McSharry, 2013), although in some applications

linear feature transformation techniques may operate

well and lead also to interpretable embedded (derived)

features where the computed latent variables may be

interpretable (van der Maaten et al., 2008a; Tsanas et

al., 2017).

Here, we explored both feature selection and

feature transformation approaches. Specifically, we

applied Principal Component Analysis (PCA) and

Independent Component Analysis (ICA), two

commonly used feature transformation methods

which aim to project the original data onto a new

feature space, which might lead to better prediction

performance (see van der Maaten et al., 2008a for

details). Whereas in PCA the resulting components

are ranked in terms of explaining the variance in the

dataset, in ICA there is no direct ranking that we

could use to understand which components should be

selected first. Therefore, the transformed features that

were computed using ICA were fed into the feature

selection algorithms (described in the next paragraph)

to decide on the transformed features to be presented

into the statistical learner. For ICA we used the

fastICA implementation.

For feature selection, we used (1) GSO; (2)

LOGO (Sun et al., 2010), a feature weighting

algorithm which implicitly also provides an estimate

of the “importance” of each feature to obtain the

ranked features; (3) minimal Redundancy Maximal

Relevance (mRMR) (Peng et al., 2005); (4) L1-LSMI

(Jitkrittum et al., 2013), and (5) SPECCMI (. In all

cases we aimed to process the top-50 selected features

from each of the algorithms. We remark that we used

GSO in the original study (Arora, Baghai-Ravary and

Tsanas, 2019) so here wanted to experiment with

different feature selection algorithms to explore

whether they bring any performance improvement.

Feature were selected using 90% of the data and

finally applying a feature selection voting strategy as

described in previous studies (Tsanas, 2012; Tsanas

et al., 2014a). We aimed to use diverse feature

selection algorithms which have been used in

different applications both to assess how stable

findings across the different feature selection

algorithms are, and also to determine whether any of

these lead to better overall classifier performance (see

the following section).

3.4 Statistical Exploration and

Mapping

We explored the statistical associations in the dataset

using standard Spearman correlation coefficients,

considering a relationship statistically strong if the

magnitude of the correlation coefficient was at least

0.3, following standard recommendation in the

medical field (Tsanas et al., 2013). This was towards

exploring both the original features and also the

transformed features from PCA and ICA to determine

whether the transformation has led to substantial

improvement in terms of feature association with the

response.

Subsequently, we used a Random Forest (RF)

algorithm, which is known to be very robust and has

been described as ‘best off-the-shelf’ algorithm for

statistical learning (Hastie, Tibshirani, Friedman,

2009). We used the default parameters (500 trees, the

number of features over which to search for the

optimal split was the square root of the number of

features, and in the end used majority voting to

determine the RF output).

3.5 Model Validation

Given the dataset is highly unbalanced ( >80%

samples belong to the dominant class, control

participants) a setting which is known to be

particularly challenging for statistical learning

models (Hastie, Tibshirani, Friedman, 2009), we

wanted to focus on a balanced dataset to avoid class

dominance problems. Specifically, we randomly

selected 1500 samples from PwP and 1500 samples

from controls to create a balanced binary

classification dataset (n=3000 samples) which will be

used to select features (or transform features), and

train the RF. We used the selected feature subset

applying standard 10-fold Cross Validation (CV) to

empirically compare performance as a function of the

number of features presented into RF.

SERPICO 2022 - Special Session on Diagnostic, Prognostic, and Phenotyping Models from Mined Administrative Healthcare Data

330

Table 2: Summary of selected features in descending order for each of the feature selection algorithms.

GSO LOGO mRMR L1-LSMI SPECCMI

Jitter->F0_TKEO_prc25 det_entropy_log_6_coef Jitter->F0_TKEO_prc95 DFA det_LT_TKEO_mean_7_coef

MFCC_4th coef MFCC_2nd coef 6th delta 6th delta det_LT_TKEO_std_7_coef

6th delta MFCC_4th coef Jitter->pitch_TKEO_prc5 det_entropy_log_6_coef det_LT_entropy_shannon_7_coef

MFCC_0th coef Jitter->F0_TKEO_prc95 VFER->SNR_SEO det_LT_entropy_log_3_coef Jitter->F0_TKEO_prc5

det_entropy_log_6_coef MFCC_0th coef VFER->std MFCC_2nd coef det_TKEO_mean_7_coef

app_TKEO_std_2_coef 6th delta DFA app_LT_entropy_log_3_coef app_entropy_log_5_coef

MFCC_9th coef app_LT_entropy_log_4_coef Jitter->pitch_PQ5_classical_Baken MFCC_5th coef app_LT_entropy_log_1_coef

IMF->NSR_TKEO app_LT_entropy_log_5_coef IMF->NSR_TKEO MFCC_4th coef det_LT_entropy_shannon_6_coef

Jitter->pitch_TKEO_prc25 MFCC_6th coef Jitter->F0_TKEO_prc5 app_LT_entropy_log_2_coef det_entropy_shannon_6_coef

MFCC_6th coef app_LT_entropy_log_3_coef MFCC_9th coef IMF->NSR_TKEO det_entropy_log_6_coef

For brevity we only present the top-10 selected features using the feature selection algorithms. For further explanation on these dysphonia

measures we refer to Tsanas (2012) and the associated toolbox freely available from https://www.darth-group.com/software and also from

https://github.com/ThanasisTsanas/VoiceAnalysisToolbox).

4 RESULTS

We started analysis by computing the correlation

coefficients of the original features. Overall, the

highest correlation coefficient was 0.14, which

already indicates this is a challenging binary

classification task. Next we computed the

transformed features using PCA and ICA and

computed the correlation coefficients: we found that

there was some minor improvement with a few more

variables exhibiting correlation coefficients with a

magnitude over 0.1, however again the highest

correlation coefficient we obtained was 0.16.

Then, we applied the feature selection algorithms

to determine the top-50 features for each algorithm.

Results are summarized in Table 2, where for brevity

we only included the top-10 features for the five feature

selection algorithms. We remark that the feature sets

obtained are quite different, although some of the

MFCCs appear to be consistently selected indicating

this is an algorithmic family that contributes to the

binary differentiation task. Similarly, many of the

wavelet features appear regularly across the feature

selection algorithms, which suggests this generic

approach of quantifying signal properties is also well-

suited to differentiating PwP from controls.

Next, we present in Figure 1 the out of sample

performance as a function of the number of features

presented into the RF for the feature selection

algorithms. This enables the exploration of different

combinations and also towards identifying a

parsimonious model where the inclusion of additional

features is not contributing to improving the model

Figure 1: Out of sample performance as a function of the

presented features into the RF, for each of the five feature

selection algorithms.

performance (or indeed leads to performance

degradation).

The results in Figure 1 suggest that we can

differentiate PwP from controls with 67.5% accuracy

using 35 features selected using either mRMR or L1-

LSMI. We remark that L1-LSMI generally performs

very well in this dataset, whereas SPECCMI clearly

underperforms by comparison.

When we tried using the transformed features into

the RF classifier the best performance obtained was

66.5%, so for this particular dataset it appears that

feature transformation has not provided any

additional benefits to improve performance.

Exploring Feature Selection and Feature Transformation Techniques to Improve Telephone-based Biomedical Speech Signal Processing

towards Parkinson’s Assessment

331

5 DISCUSSION

We investigated the potential of differentiating PwP

and controls using telephone-recorded speech

collected under acoustically non-controlled

conditions exploring different feature selection and

feature transformation methods. We found that the

most frequently used feature transformation methods,

PCA and ICA do not appear to provide any

improvement in the classification accuracy compared

to the investigated feature selection approaches.

Overall, we found that we can differentiate the two

groups with about 67% accuracy, which improves on

(Arora, Baghai-Ravary and Tsanas, 2019).

Compared to the earlier study (Arora, Baghai-

Ravary and Tsanas, 2019) which pooled together all

the available data in PVI, here we focused only on a

balanced subset of 3000 participants from the Boston

cohort in PVI. The underlying reason is that focusing

on participants coming from the same linguistic

background, even when only processing sustained

vowel /a/ phonations, would mitigate potential

differences. Moreover, by selecting a balanced subset

of the data we overcome the common challenging

setting where the dominant class may skew the

classifier’s outputs. This has indeed led to some

performance improvement (previously the best

performing model in (Arora, Baghai-Ravary and

Tsanas, 2019) led to 63.7% balanced accuracy,

whereas here we report 67% accuracy (which by

definition coincides with the balanced accuracy given

we have a balanced dataset).

We found that although the feature transformation

methods explored herein (PCA and ICA) led to some

transformed features that univariately were slightly

better correlated with the response compared to the

original features, when taken jointly they did not lead

to better classification outcomes. Therefore, we did

not pursue this further since feature transformation

methods also have the disadvantage that the resulting

models are less interpretable. It is possible that some

more convoluted feature transformation methods

(e.g. see van der Maaten et al., 2008a) might perform

better here, and this is an area that needs to be

explored in further work. Also, we did not explore

further data visualization approaches to explore

projected feature subsets, which may provide

tentative insights into the differences of samples

between classes (van der Maaten et al., 2008b).

Previous work that used the entire PVI dataset

(Arora, Baghai-Ravary and Tsanas, 2019) and GSO

to determine the best performing feature subset using

the same methodology as explored in this study led to

quite different features. This likely supports earlier

findings that even for sustained vowels there may be

subtle differences given the linguistic background of

participants. In turn, this has important implications

towards developing generalizable tools across

cohorts of participants coming from different

linguistic backgrounds.

We found that substantially different feature

subsets (using mRMR and L1-LSMI) lead to very

similar performance in the RF. This likely indicates

the presence of different Markov blankets in the

dataset, where quite different features lead to similar

out of sample performance. This is in accordance to

previous findings in this field with different speech-

PD datasets (e.g. see Tsanas 2012) and possibly

underlines the fact there may be different underlying

combinations of features which essentially can jointly

capture the key acoustic characteristics towards

differentiating PwP from controls.

We remark that although the reported

performance is comparably low to apply this tool in

clinical practice currently, it is possible that it could

be used as an early indicator, particularly given there

is practically no cost to deploy the use of sustained

vowels in practice and collect data through standard

telephone networks. It is likely that in combination

with additional signal modalities (e.g. walking) and

other tests that can be collected using smartphones

(e.g., see Tsanas et al., 2020; Woodward et al., in

press), we will be able to develop an affordable and

practical tool to change contemporary PD screening

and facilitate early diagnosis.

Collectively, this study’s findings are a step

towards developing a robust, effective and cost-

efficient tool to screen for PD at large.

ACKNOWLEDGEMENTS

We are grateful to Max Little who led the Parkinson’s

Voice Initiative where the data for this study was

collected, and to Ladan Baghai-Ravary for

developing the data collection process using the

Aculab servers. We would like to extend our thanks

to all participants in the PVI study. The study was

made possible through generous funding via an

EPSRC-NCSML award to AT and SA.

REFERENCES

Arora, S. Visanji, N.P., Mestre, T.A., Tsanas, A., Al

Dakheel, A., Connolly, B.S., Gasca-Salas, C., Kern,

D.S., Jain, J., Slow, E.J., Faust-Socher, A., Lang, A.E.

Little, M.A., Marras C. 2018. Investigating voice as a

SERPICO 2022 - Special Session on Diagnostic, Prognostic, and Phenotyping Models from Mined Administrative Healthcare Data

332

biomarker for leucine-rich repeat kinase 2-associated

Parkinson’s disease: a pilot study, Journal of

Parkinson’s Disease, Vol. 8(4), pp. 503-510

Arora, S., Baghai-Ravary, L., Tsanas A. 2019. Developing

a large scale population screening tool for the

assessment of Parkinson’s disease using telephone-

quality speech, Journal of Acoustical Society of

America, Vol. 145(5), 2871-2884

Arora, S., Tsanas A. 2021a. Assessing Parkinson’s Disease

at Scale using Telephone-recorded Speech: Insights

from the Parkinson’s Voice Initiative, Diagnostics, Vol.

11:1892

Arora, S., Lo, C., Hu, M., Tsanas A. 2021b. Smartphone

speech testing for symptom assessment in rapid eye

movement sleep behavior disorder and Parkinson’s

disease, IEEE Access, Vol. 9, pp. 44813-44824

Bishop, C.M. 2006. Pattern recognition and machine

learning, Springer

Camacho, A., Harris, J.G. 2008. A sawtooth waveform

inspired pitch estimator for speech and music, Journal

of the Acoustical Society of America, Vol. 124, 1638-

1652

Dorsey, E.R., George, B.P., Leff, B., Willis A.W. 2013. The

coming crisis: obtaining care for the growing burden of

neurodegenerative conditions, Neurology, Vol. 80,

1989-1996

Duda, R.O., Hart, P.E., Stork, D.G. 2001. Pattern

classification, Wiley-interscience, 2nd ed.

GBD 2016 Parkinson's Disease Collaborators 2018. Global,

regional, and national burden of Parkinson’s disease,

1990–2016: a systematic analysis for the Global Burden

of Disease Study 2016. The Lancet Neurology, Vol. 17,

pp. 939-953

GBD Neurological Disorders Collaborators 2021. Burden

of Neurological Disorders Across the US From 1990-

2017: A Global Burden of Disease Study. JAMA

Neurology, Vol. 78(2), pp. 165-176

Gomez-Vilda, P., Mykyska, J., Gomez, A., Palacios, D.,

Rodellar, V., Alvarez A. 2019. Characterization of

Parkinson’s disease dysarthria in terms of speech

articulation kinematics, Biomedical Signal Processing

and Control, Vol. 52, 312-320

Gorriz, J.M., Ramirez, J., Ortiz, A., et al. (2020). Artificial

intelligence within the interplay between natural and

artificial computation: advances in data science, trends

and applications, Neurocomputing, Vol. 410, pp. 237-

270

Guyon, I., Gunn, S., Nikravesh, M., Zadeh, L.A. (Eds.)

2006. Feature Extraction: Foundations and

Applications, Springer

Hartelius L., Svensson P. 1994. Speech and swallowing

symptoms associated with parkinson‘s disease and

multiple sclerosis: A survey, Folia Phoniatr. Logop.,

Vol. 46, pp. 9- 17

Hastie, T. Tibshirani, R. Friedman J. 2009. The elements of

statistical learning: data mining, inference, and

prediction, Springer, 2nd ed.

Horne, E., Tibble, H. Sheikh, A., Tsanas A. 2020.

Challenges of clustering multimodal clinical data: a

review of applications in asthma subtyping, JMIR

Medical Informatics, Vol. 8(5), e16452

Jitkrittum, W., Hachiya, H. and Sugiyama, M. (2013)

Feature selection via L1-penalized squared-loss mutual

information, IEICE Transactions on Information and

Systems, E96-D(7), pp. 1513–1524

Lewis, S.J.G., Foltynie, T., Blackwell, A.D., Robbins,

T.W., Owen, A.m., Barker R.A. 2005. Heterogeneity of

Parkinson’s disease in the early clinical stages using a

data driven approach, Journal of Neurology,

Neurosurgery and Psychiatry, Vol. 76, 343-348

Mu, J., Chaudhuri, K.R., Bielza, C., de Pedro-Cuesta, J.,

Larranaga, P., Martinez-Martin, P. 2017. Parkinson’s

disease subtypes identified from cluster analysis of

motor and non-motor symptoms, Frontiers in Aging

Neuroscience, 9:301

Naylor, P.A., Kounoudes, A., Gudnason J., Brookes, M.

(2007). Estimation of glottal closure instants in voiced

speech using the DYPSA algorithm, IEEE Transactions

on Audio, Speech, and Language Processing, Vol.

15(1), pp. 34-43

Olanow, C.W., Stern, M.B., Sethi,K. 2009. The scientific

and clinical basis forthe treatment of Parkinson disease,

Neurology, Vol. 72 (21 Suppl 4) s1-s136

Peng, H., Long., F., Ding, C. 2005. Feature Selection Based

on Mutual Information: Criteria of Max-Dependency,

Max-Relevance, and Min-Redundancy, IEEE

Transactions on pattern analysis and machine

intelligence, Vol. 27, pp. 1226-1238

Roark, R.M. 2006. Frequency and Voice: Perspectives in

the Time Domain, Journal of Voice, Vol. 20(3), pp.

325-354

Rueda, A., Krishnan, S., 2018. Clustering Parkinson’s and

age-related voice impairment signal features for

unsupervised learning, Advances in Data Science and

Adaptive Analysis, Vol. 10(2);1840007

San Segundo, E., Tsanas, A., Gomez-Vilda, P., 2017.

Euclidean distances as measures of speaker similarity

including identical twin pairs: a forensic investigation

using source and filter voice characteristics, Forensic

Science International, Vol. 270, pp.25-38

Sheaves, B., Porcheret, K., Tsanas, A., Espie, C., Foster, R.,

Freeman, D., Harrison, P.J., Wulff, K., Goodwin, G.M.

2016. Insomnia, nightmares, and chronotype as markers

of risk for severe mental illness: results from a student

population, Sleep, Vol. 39(1), pp. 173-181

Sun, Y., Todorovic, S., Goodison, S. 2010. Local-learning-

based feature selection for high-dimensional data

analysis, IEEE Transactions on Pattern Analysis and

Machine Intelligence, Vol. 32(9), pp. 1610-1626

Titze, I.R. 2000. Principles of Voice Production. National

Center for Voice and Speech, Iowa City, US, 2nd

printing

Triantafyllidis, A.K., Tsanas A. 2019. Applications of

machine learning in real-life digital health

interventions: review of the literature, Journal of

Medical Internet Research (JMIR), Vol. 21(4), e12286

Tsanas, A., Little, M.A., McSharry, P.E., Ramig, L.O.

2010a. New nonlinear markers and insights into speech

signal degradation for effective tracking of Parkinson’s

Exploring Feature Selection and Feature Transformation Techniques to Improve Telephone-based Biomedical Speech Signal Processing

towards Parkinson’s Assessment

333

disease symptom severity, International Symposium on

Nonlinear Theory and its Applications (NOLTA), pp.

457-460, Krakow, Poland, 5-8 September

Tsanas, A., Little, M.A., McSharry, P.E., Ramig, L.O.

2010b. Enhanced classical dysphonia measures and

sparse regression for telemonitoring of Parkinson's

disease progression, IEEE Signal Processing Society,

International Conference on Acoustics, Speech and

Signal Processing (ICASSP), pp. 594-597, Dallas,

Texas, US, 14-19 March 2010

Tsanas, A., Little, M.A., McSharry, P.E., Ramig, L.O.

2010c. Accurate telemonitoring of Parkinson’s disease

progression by non-invasive speech tests, IEEE

Transactions on Biomedical Engineering, Vol. 57, pp.

884-893

Tsanas, A., Little, M.A., McSharry, P.E., Ramig, L.O.

2011. Nonlinear speech analysis algorithms mapped to

a standard metric achieve clinically useful

quantification of average Parkinson’s disease symptom

severity, Journal of the Royal Society Interface, Vol. 8,

pp. 842-855

Tsanas, A., Little, M.A., McSharry, P.E., Spielman, J.,

Ramig, L.O. 2012. Novel speech signal processing

algorithms for high-accuracy classification of

Parkinson’s disease, IEEE Transactions on Biomedical

Engineering, Vol. 59, 1264-1271

Tsanas A. 2012. Accurate telemonitoring of Parkinson’s

disease symptom severity using nonlinear speech signal

processing and statistical machine learning, Ph.D.

thesis, Oxford Centre for Industrial and Applied

Mathematics, University of Oxford

Tsanas A. 2013. Acoustic analysis toolkit for biomedical

speech signal processing: concepts and algorithms, 8th

International Workshop on Models and Analysis of

Vocal Emissions for Biomedical Applications

(MAVEBA), pp. 37-40, Florence, Italy, 16-18

December

Tsanas, A., Gómez-Vilda P., 2013. Novel robust decision

support tool assisting early diagnosis of pathological

voices using acoustic analysis of sustained vowels,

Multidisciplinary Conference of Users of Voice, Speech

and Singing (JVHC 13), pp. 3-12, Las Palmas de Gran

Canaria, 27-28 June

Tsanas, A. Little, M.A., McSharry P.E. 2013. A

methodology for the analysis of medical data, in

Handbook of Systems and Complexity in Health, Eds.

J.P. Sturmberg, and C.M. Martin, Springer, pp. 113-125

(chapter 7)

Tsanas, A., Little, M.A., Fox, C., Ramig L.O. 2014a.

Objective automatic assessment of rehabilitative speech

treatment in Parkinson’s disease, IEEE Transactions on

Neural Systems and Rehabilitation Engineering, Vol.

22, 181-190

Tsanas, A., Zañartu, M., Little, M.A., Fox, C., Ramig, L.O.,

Clifford, G.D. 2014b. Robust fundamental frequency

estimation in sustained vowels: detailed algorithmic

comparisons and information fusion with adaptive

Kalman filtering, Journal of the Acoustical Society of

America, Vol. 135, 2885-2901

Tsanas, A., Saunders, K.E.A., Bilderbeck, A.C., Palmius,

N., Goodwin, G.M., De Vos, M. 2017. Clinical insight

into latent variables of psychiatric questionnaires for

mood symptom self-assessment, JMIR Mental Health,

Vol. 4, No. 2, pp. e15

Tsanas A. 2019. New insights into Parkinson’s disease

through statistical analysis of standard clinical scales

quantifying symptom severity, 41st IEEE Engineering

in Medicine in Biology Conference (EMBC), Berlin,

Germany, 23-27 July

Tsanas, A., Arora S. 2019. Biomedical speech signal

insights from a large scale cohort across seven

countries: the Parkinson’s voice initiative study,

Models and Analysis of Vocal Emissions for Biomedical

Applications, Florence, Italy, 17-19 December 2019

Tsanas, A., Arora S. 2020. Large-scale clustering of people

diagnosed with Parkinson’s disease using acoustic

analysis of sustained vowels: findings in the

Parkinson’s voice initiative study, 13th International

Joint Conference on Biomedical Systems and

Technology (BIOSTEC), pp. 369-376, Malta, 26-28

February 2020

Tsanas, A., Woodward, E., Ehlers, A. 2020. Objective

characterization of activity, sleep, and circadian rhythm

patterns using a wrist worn sensor: insights into post-

traumatic stress disorder, JMIR mHealth and uHealth,

Vol. 8(4), pp. e14306

Tsanas, A., Little, M.A., Ramig L.O. 2021. Remote

assessment of Parkinson’s disease symptom severity

using the simulated cellular mobile telephone network,

IEEE Access, Vol. 9, pp. 11024-11036

Tsanas, A., Arora S. 2021a. in the Parkinson’s voice

initiative study, 14th International Joint Conference on

Biomedical Systems and Technology (BIOSTEC), pp.

124-131, Vienna, Austria, 11-13 February 2021

Tsanas, A. Arora S. 2021b. Acoustic analysis of sustained

vowels in Parkinson’s disease: new insights into the

differences of UK- and US-English speaking

participants from the Parkinson’s voice initiative,

MAVEBA, Florence, Italy, 14-16 December

Woodward, K. Kanjo, E., Brown, D., et al. (2020). Beyond

mobile apps: a survey of technologies for mental well-

being, IEEE Transactions Affective Computing, (in

press)

van der Maaten, L.P.J., Postma, E.O., van den Herik, H.J.

2008a. Dimensionality reduction: a comparative

review, Technical Report, University of Tilburg

van der Maaten, L.J.P., Hinton, G.E. 2008b. Visualizing

high-dimensional data using t-SNE, Journal of

Machine Learning Research , Vol. 9, pp. 2579-2605

Vinh, X. N., Chan, J., Romano, S., Bailey, J. (2014)

Effective global approaches for mutual information

based feature selection, in Proceedings of the ACM

SIGKDD International Conference on Knowledge

Discovery and Data Mining, pp. 512–521. doi:

10.1145/2623330.2623611

SERPICO 2022 - Special Session on Diagnostic, Prognostic, and Phenotyping Models from Mined Administrative Healthcare Data

334