Feature Selection Improves Speech Based Parkinson’s Disease Detection

Performance

Ays¸e Nur Tekindor

1 a

and Eda Akman Aydın

2 b

Electrical and Electronics Engineering, Graduate School of Natural and Applied Sciences,

Gazi University, Ankara, Turkey

Department of Electrical and Electronics Engineering, Faculty of Technology, Gazi University, Ankara, Turkey

Keywords:

Parkinson’s Disease, Speech Analysis, mRMR, Bark Spectrum, Mel Frequency Cepstral Coefﬁcients,

Gammatone Cepstral Coefﬁcients, Spectral Features.

Abstract:

Parkinson’s disease (PD) is a neurodegenerative disorder that is caused by decrease in dopamine levels in

the brain. There is currently no cure for PD; however, the progression of the disease can be brought under

control by diagnosis made in early stages. Studies have shown that speech impairments are early symptoms

of PD. In this study, an approach for the early diagnosis of patients with PD using speech based features was

proposed. In order to detect the PD, four feature groups such as Bark Spectrum coefﬁcients, Mel Frequency

Cepstral Coefﬁcients (MFCCs), Gammatone Cepstral Coefﬁcients (GTCCs), and Spectral-Temporal Features

were created. Minimum Redundancy Maximum Relevance (mRMR) based feature selection was applied

to each feature group. Three classiﬁers including decision tree, Naive Bayes, and support vector machine

were employed to evaluate the performance of the feature sets. The proposed method was validated on the

Italian speech dataset. Feature selection improved the PD diagnosing performance, especially for the Naive

Bayes model which obtained 96.01% accuracy by overall feature selection and 96.17% by group-based feature

selection.

1 INTRODUCTION

Parkinson’s disease (PD) is a neurodegenerative dis-

ease that is the result of decreased dopamine level,

which is a brain chemical produced by neurons work-

ing as a neurotransmitter in the brain (Appakaya and

Sankar, 2018). The reason behind PD is still un-

known, but genetic and environmental factors are

thought to be the cause ((Polat and Nour, 2020),

(Tolosa et al., 2021)). There is no deﬁnitive treatment

for the disease; however, some drugs used to control

symptoms in the early stages have an important effect

on the progression of the disease. Changes in speech

and handwriting, tremors, slowed movements, muscle

stiffness, deterioration of postural and balance control

and loss of automatic movements are common symp-

toms in Parkinson’s patients.

Vocal problems are one of the most important

symptoms seen in the early stages of the disease in

approximately 90% of people with PD (Connolly and

Lang, 2014). The number of studies focused on de-

https://orcid.org/0000-0002-6124-5621

https://orcid.org/0000-0002-9887-3808

tecting PD from speech signals in early stages of the

disease is increasing day by day. Dataset collected by

(Sakar et al., 2013), which is one of the ﬁrst examples

in the literature, has been used in many studies. This

dataset consists of extracted features including time-

frequency-based features of audio signals of healthy

controls (HC) and patients with PD. In a study (Priya

et al., 2021) that uses this dataset; decision tree, naive

bayes, support vector machines, k-nearest neighbor,

random forest classiﬁers were compared obtaining ac-

curacy rates of 91, 61, 64, 64 and 95% respectively.

Since the dataset consists of extracted features instead

of original audio signals, studies using this dataset

do not include preprocessing and feature extraction

steps. In another study, (Dimauro and Girardi, 2019)

published the dataset they collected to use in their

study (Dimauro et al., 2017) where a system designed

to convert the voice of PD patients into text. This

dataset contains the original speech signals, and it was

used in several studies to examine feature extraction

techniques.

A study (Appakaya et al., 2020) that uses the

same dataset by extracting Mel Frequency Cepstral

726

Tekindor, A. and Aydın, E.

Feature Selection Improves Speech Based Parkinson’s Disease Detection Performance.

DOI: 10.5220/0012347300003657

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 1, pages 726-732

ISBN: 978-989-758-688-0; ISSN: 2184-4305

Coefﬁcients (MFCC) and pitch synchronous features

trained 25 classiﬁers and obtained a mean classiﬁca-

tion accuracy of 88.5%. Another study (Appakaya

et al., 2021) that used the same dataset achieved 85%

classiﬁcation accuracy with Logistic regression and

Linear SVM using the features extracted by autoen-

coders. With the application of leave-one-subject-out

(LOSO) classiﬁcation, 84% accuracy was obtained.

(Lamba et al., 2023) used a hybrid dataset, one of

them is the same Italian speech dataset. The ex-

tracted features include duration, fundamental fre-

quency, harmonic to noise ratio (HNR), jitter, shim-

mer, and principal component analysis (PCA). After

applying a genetic algorithm method for feature se-

lection, k-nearest neighbor, XGBoost, random forest,

and logistic regression classiﬁers obtained 90% mean

accuracy.

In this study, it is aimed to evaluate the effect of

feature extraction and feature selection on the per-

formance of Parkinson’s disease diagnosis based on

speech signals. In this scope, four feature groups,

such as, Bark Spectrum coefﬁcients, Mel Frequency

Cepstral Coefﬁcients (MFCCs), Gammatone Cepstral

Coefﬁcients (GTCCs) and other (spectral-temporal)

features were extracted. The most signiﬁcant features

in the groups were determined by the feature selec-

tion method based on Minimum Redundancy Max-

imum Relevance (mRMR). Proposed approach was

validated on the dataset collected by (Dimauro and

Girardi, 2019). The performance of the proposed

method was evaluated by accuracy, precision, recall,

F1 score, and feature reduction rate metrics. The

results of the study revealed that the mRMR-based

feature selection method could improve the accuracy

rates.

This study consists of ﬁve sections. In Section 2,

the dataset, feature extraction methods, feature selec-

tion and classiﬁcation method used in the study are

explained. Experimental results are presented in Sec-

tion3. Theresultsofthestudyareinterpreted in Section

4, and ﬁnally, the study is summarized in Section 5.

2 METHODOLOGY

2.1 Dataset

The dataset used in this study was collected by (Di-

mauro et al., 2017) and can be accessed from IEEE

Dataport. Italian speech recordings of 50 speakers

with 29 males (19 PD, 10 HC) and 21 females (9

PD, 12 HC) are contained. Speakers with PD aged

between 40 and 80 years and healthy controls aged

between 60 and 77 years. Original dataset consists

of records where a passage that is phonetically bal-

anced is read twice, /pa/ and /ka/ syllables for 5 sec,

reading of some phonemically balanced words, read-

ing of some phonemically balanced phrases and two

phonations of the vowels /a/, /e/, /i/, /o/ and /u/. Sam-

pling frequency for speech samples is 16 kHz. Dis-

tance between the speaker and microphone is 15 to

25 cm and recordings were made in noise-free con-

ditions. Speakers exhibit no speech pathology except

for PD. In the scope of this study, exclusive utiliza-

tion was made of vowel recordings from the dataset

encompassing speech signals. Two speakers of HC

whose recordings are poor in quality were excluded.

Also, after some trials of feature extraction, it was

observed that the dataset is dramatically imbalanced

in terms of recording length. In order to handle this

problem, speakers with a long recording time were

excluded and PD-HC distribution was balanced. One

vowel phonation of other speakers for each group was

included.

2.2 Feature Extraction

The datasets contain voice recordings stored as

“.wav” ﬁles. MATLAB is utilized for both fea-

ture extraction and classiﬁcation. Speech signals ex-

hibit non-stationarity due to variations in their statis-

tical characteristics over time. Applying windowing

techniques to process speech signals involves treat-

ing them as approximately stationary (Kumar et al.,

2022). Typically, window durations fall within the

range of 20 to 40 ms. In this study, we divided the

dataset into non-overlapping segments of 50 ms each,

resulting in 800 samples per window. 73 features

in four feature groups were extracted. They include

bark spectrum, MFCC, GTCC, and other features

consisting of spectral centroid, spectral crest, spectral

decrease, spectral entropy, spectral ﬂatness, spectral

ﬂux, spectral kurtosis, spectral rolloff point, spectral

skewness, spectral slope, spectral spread, consisting

of pitch, harmonic ratio, zero cross rate, and short

time energy. The descriptions of the extracted fea-

tures are given below:

2.2.1 Bark Spectrum

Bark spectrum refers to a psychoacoustic frequency

scale that approximates the human perception of

sound.

2.2.2 Mel Frequency Cepstral Coefﬁcients

(MFCC)

MFCC is the expression of the short-time power spec-

trum of the audio signal on the Mel scale.

Feature Selection Improves Speech Based Parkinson’s Disease Detection Performance

727

2.2.3 Gammatone Cepstral Coefﬁcients

(GTCCs)

GTCCs are features derived from Gammatone ﬁlters

which are inspired biologically to approximate the au-

ditory ﬁltering performed by the human ear.

2.2.4 Other Features

This feature group includes other spectral and tempo-

ral features.

Spectral Centroid: Spectral centroid is a measure

that indicates the central frequency or “balance point”

of a spectrum.

Spectral Crest: Spectral crest measures the relation-

ship between the highest point in a spectrum and the

average value of the spectrum.

Spectral Decrease: Spectral decrease measures the

extent of the spectrum’s decrease, focusing on the

slopes of lower frequencies.

Spectral Entropy: Spectral entropy gauges the peak-

iness of the spectrum.

Spectral Flatness: Spectral ﬂatness calculates the ra-

tio between the geometric mean and the arithmetic

mean of a spectrum.

Spectral Flux: Spectral ﬂux measures the variability

of a spectrum over time.

Spectral Kurtosis: Spectral kurtosis quantiﬁes the

ﬂatness or non-Gaussianity of a spectrum around its

center frequency.

Spectral Rollof Point: Spectral rolloff point mea-

sures the bandwidth of the audio signal by determin-

ing the frequency bin under which a given percentage

of the total energy exists.

Spectral Skewness: Spectral skewness measures the

symmetry around the spectral centroid and is also

known as spectral tilt in phonetics. It, along with

other spectral moments helps distinguish the place of

articulation.

Spectral Slope: Spectral slope quantiﬁes the de-

crease in a spectrum.

Spectral Spread: Spectral spread calculates the stan-

dard deviation around the spectral centroid and repre-

sents the “instantaneous bandwidth” of a spectrum.

Pitch: Pitch is used to describe the height or lowness

of a tone as perceived by the ear.

Harmonic Ratio: Harmonic ratio refers to a measure

that quantiﬁes the presence and strength of harmonics

in a sound signal.

Zero Cross Rate: Zero-crossing rate is a measure

that quantiﬁes the rate at which a signal changes its

sign (from positive to negative or vice versa) over

time.

Short Time Energy: Short time energy speciﬁes

the signal amplitude of a certain signal point over

a period of time ((Kumar et al., 2022),(Hawi et al.,

2022),(Boualoulou et al., 2023), (Priya et al., 2022),

(Majda-Zdancewicz et al., 2022), (Hassan et al.,

2022), (Chu et al., 2009)).

2.3 Feature Selection

Feature selection is an essential step since it deter-

mines the most distinguishing features and reduces

computational time. One of the objectives of our

study is to indicate the effect of feature selection

on classiﬁcation performance. In this context, the

mRMR method was used.

Minimum Redundancy Maximum Relevance

(mRMR): MRMR algorithm is a widely recognized

technique for selecting relevant features in a feature

set. It accomplishes this by evaluating both the

redundancy between features, and the relevance

between features and the target variable. To assess

redundancy and relevance, mRMR employs the

concept of mutual information from information

theory. mRMR ranks features by their predictive

signiﬁcance concerning the target variable, taking

into account both relevance and redundancy. This

ranking allows for identifying important features for

accurate predictions while minimizing redundant

information among them (Radovic et al., 2017).

Two scenarios were studied where the mRMR al-

gorithm was applied, as shown in Figure 2. Firstly,

the top 25 features (most important ones) of the train-

ing data were selected from the entire feature set by

mRMR. Secondly, each single feature group (Bark,

MFCC, GTCC, Other) was evaluated alone in terms

of feature importance. Subsets were created by select-

ing the top 9 features for bark group, top 4 features for

MFCC group, top 4 features for GTCC group, and top

8 features for other group. Then, these subsets were

combined.

2.4 Classiﬁcation

After feature extraction and feature selection by

mRMR method, different classiﬁcation tasks were

done. Dataset were split into train and test sets with

the ratio of 80% and 20% respectively. 10-fold cross-

validation was applied in each session. Three dif-

ferent conditions were observed during the experi-

ment. Firstly, three classiﬁers including decision tree,

Naive Bayes and support vector machine models were

trained by all extracted features. Then, overall se-

lected features and combination of group-based se-

lected features were used to train the same classiﬁers

respectively. The evaluation metrics include accu-

racy, precision, recall, F1-score and Feature Reduc-

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

728

Figure 1: Experimental procedure of the proposed method.

tion Rate score. The formulations of these metrics are

given in Equations 1-4.

Accuracy =

T P + T N

T P + FP + T N + FN

(1)

Precision =

T P

T P + FP

(2)

Recall =

T P

T P + FN

(3)

F1 − Score =

2 ∗ Precision ∗ Recall

Precision + Recall

(4)

where TP, TN, FP, FN are the numbers of true pos-

itives, true negatives, false positives and false nega-

tives. Accuracy is the percentage of data points cor-

rectly predicted out of the total data availability. Pre-

cision is the accuracy of positive predictions. Recall

is the measure of the model correctly identifying true

positives. F1- Score is the harmonic mean of preci-

sion and recall. Evaluating accuracy and F1 score to-

gether gives more meaningful results.

Feature Reduction Rate (FRR) is the ratio of the

number of removed features to the original number

of features. It is computed using Equation 5 (Aydin,

2020):

FRR = 1 −

Number of selected features

Number of all features

(5)

Number of the selected features and FRR for over-

all feature selection and group-based feature selection

are given in Table 1. In Figure 1, experimental proce-

dure of the proposed approach is shown step by step.

3 RESULTS

In the step of feature extraction, 32 in bark spectral

features group (group 1), 13 in MFCC features group

(group 2), 13 in GTCC features group (group 3) and

15 in other features group (group 4); 73 features in to-

tal were extracted. Feature groups and number of fea-

tures are shown in Table 1. MRMR algorithm was ap-

plied for feature selection. Feature importance graphs

for each feature group are shown in Figure 2.

Decision tree, Naive Bayes and SVM models

were chosen as classiﬁers since they show high per-

formance on disease detection from speech signals.

In this study, experiments were carried out in three

stages. In the ﬁrst stage, the performance of the

models in diagnosing Parkinson’s disease was exam-

ined by using all the features in the four extracted

feature groups. In the second stage, evaluation was

made using the overall subset, where features were se-

lected from the entire feature set. Finally, evaluation

was made using the combination subset where group-

based extracted features were combined. Classiﬁer

parameters were speciﬁcally determined for each ex-

periment. 10-fold cross-validation was applied in the

validation of models. The results are given in Ta-

ble 2. Both overall subset and combination subset

contain 25 features. This ensures that the feature re-

duction rate was 65.75For all features, the decision

tree model achieved an accuracy of 93.83%, with a

high recall of 95.94%. This indicates that the model

successfully identiﬁed the majority of true positive

cases (PD). 92.73% precision and 94.31% F1 score

were also commendable, suggesting a balanced trade-

off between precision and recall. The Naive Bayes

model achieved 89.69% accuracy, slightly lower than

the Decision Tree. However, it showed a remark-

able recall of 96.88%, indicating a strong ability to

correctly classify positive cases. 90.93% F1 score

showed a reasonable balance between precision and

recall. SVM achieved 94.37% accuracy, with a high

precision of 97.12%. It indicates the SVM model

has a strong ability to correctly classify positive in-

stances while showing a good overall performance

with 92.17% recall and 94.58% F1 score.

For overall selected features, all models demon-

strated improved performance. The Decision Tree

model achieved an accuracy of 94.20%, maintaining a

Feature Selection Improves Speech Based Parkinson’s Disease Detection Performance

729

Figure 2: Importance score of a) Bark Spectral features, b) Other features, c) MFCC features, d) GTCC features.

Table 1: Number of the selected features and Feature Re-

duction Rates.

Feature

Subset

Number

Feature

Number

Selected

Feature

(Over-

all)

Number

Selected

Feature

(Group

based)

FRR

(Over-

all -

Group

Based)

Bark

32 4 9 87.5 -

72.88

MFCC 13 5 4 61.53 -

69.23

GTCC 13 6 4 56.85 -

69.23

Other

Fea-

tures

15 10 8 33.33 -

46.67

balance between precision, recall and F1 score. Naive

Bayes showed signiﬁcant improvement, with an accu-

racy of 96.01%. It showed higher performance than

other models in the experiment with 96.59% precision

and 96.24% F1 score. SVM also improved, achieving

an accuracy of 94.92%, with a precision of 93.15%

and a high recall of 97.65%. The F1 Score for SVM

was 95.35%.

Similar to overall feature selection, group-based

feature selection improved the performance of the De-

cision Tree, Naive Bayes and SVM models. The De-

cision Tree obtained 96.36% accuracy, 95.45% preci-

sion and a high Recall of 97.84%. The F1 Score was

96.63%. Feature selection provided a balanced trade-

off between recall and other metrics for Naive Bayes

Table 2: Performance of the models according to different

feature subsets.

All Features

Model Accuracy

Precision

Recall

Score

Decision

Tree

93.83 92.73 95.94 94.31

Naive

Bayes

89.69 85.66 96.88 90.93

SVM 94.37 97.12 92.17 94.58

Selected Features (Overall Subset)

Decision

Tree

94.20 93.07 96.30 94.66

Naive

Bayes

96.01 96.59 95.89 96.24

SVM 94.92 93.15 97.65 95.35

Group-Based Selected Features

(Combination Subset)

Decision

Tree

96.36 95.45 97.84 96.63

Naive

Bayes

96.17 96.60 96.20 96.40

SVM 94.71 96.15 93.84 94.98

with 96.17% accuracy, 96.60% precision, 96.20% re-

call and 96.40% F1 score. SVM achieved an accuracy

of 94.71%, with a precision of 96.15% and a recall of

93.84%. The F1 score for SVM in this scenario was

94.98%. From a general perspective, an increase in

accuracy rates was observed. Besides, it is seen that

F1 scores are also improved with feature selection.

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

730

4 DISCUSSION

The results indicate that both overall feature selection

and group-based feature selection improved the per-

formance of the decision tree, Naive Bayes, and SVM

models. Also, group-based feature selection showed

higher performance than overall feature selection for

all models. In this study, the number of features was

reduced from 73 to 25 by feature selection, resulting

in a 65% reduction rate; at the same time, an increase

in accuracy and F1 score was observed for all classi-

ﬁers. It indicates strong overall performance.

The study using the same dataset (Appakaya

et al., 2020) obtained 88.5% accuracy. However,

this performance was obtained by applying leave-

one-subject-out (LOSO). LOSO is a cross-validation

method that indicates the performance reliability for

entirely new data. With this advantage, it can be

preferred in diagnostic decision support systems. In

this study, 10-fold cross-validation was applied dur-

ing the experiments, and 96.36% accuracy was ob-

tained. Feature selection not only increased accuracy

across all classiﬁer types, but also improved results by

enhancing other metrics closer to a balanced ratio.

The results underscore the importance of feature

engineering and model selection in achieving the best

possible classiﬁcation performance. Feature selection

excludes the features with lower importance for the

training process and increases the classiﬁcation per-

formance. By choosing the proper subset of the fea-

tures, classiﬁcation performance could be improved.

5 CONCLUSION

Parkinson’s disease (PD) is a neurodegenerative con-

dition characterized by a decrease in dopamine levels

in the brain. Currently, there is no known cure for PD,

but early diagnosis plays a crucial role in managing

the progression of the disease. In this study, an ap-

proach for early detection of PD using three feature

subsets obtained by speech analysis was proposed,

and the effect of feature selection on classiﬁcation

performance was observed. MRMR based feature se-

lection was applied to deﬁne the most discriminative

features. The study revealed a potential improvement

in the classiﬁcation performance by selecting impor-

tant features in speech-based PD diagnosis.

ACKNOWLEDGEMENTS

This work was supported by the Scientiﬁc and Tech-

nological Research Council of Turkey (T

ITAK) as

2224-A project. The authors heartedly thank and ap-

preciate T

ITAK for their support.

REFERENCES

Appakaya, S. B. and Sankar, R. (2018). Classiﬁcation of

parkinson’s disease using pitch synchronous speech

analysis. In 2018 40th Annual International Confer-

ence of the IEEE Engineering in Medicine and Biol-

ogy Society (EMBC).

Appakaya, S. B., Sankar, R., and Ra, I.-H. (2020). Classiﬁer

comparison for two distinct applications using same

data. In 9th International Conference on Smart Media

and Applications. SMA 2020: The 9th International

Conference on Smart Media and Applications.

Appakaya, S. B., Sankar, R., and Sheybani, E. (2021).

Novel unsupervised feature extraction protocol us-

ing autoencoders for connected speech: Application

in parkinson’s disease classiﬁcation. In Wireless

Telecommunications Symposium (WTS), page 1–5.

IEEE.

Aydin, E. A. (2020). Subject-speciﬁc feature selection

for near infrared spectroscopy based brain-computer

interfaces. Computer Methods and Programs in

Biomedicine, 195:105535.

Boualoulou, N., Drissi, T. B., and Nsiri, B. (2023). Cnn

and lstm for the classiﬁcation of parkinson’s disease

based on the gtcc and mfcc. Applied Computer Sci-

ence, 19(2):1–24.

Chu, S., Narayanan, S., and Kuo, C. C. J. (2009). Environ-

mental sound recognition with time–frequency audio

features. IEEE Transactions on Audio, Speech, and

Language Processing, 17(6):1142–1158.

Connolly, B. S. and Lang, A. E. (2014). Pharmacological

treatment of parkinson disease. JAMA, 311:1670.

Dimauro, G. and Girardi, F. (2019). Italian parkinson’s

voice and speech.

Dimauro, G., Nicola, V. D., Bevilacqua, V., Caivano, D.,

and Girardi, F. (2017). Assessment of speech intelli-

gibility in parkinson’s disease using a speech-to-text

system. IEEE Access, 5:22199–22208.

Hassan, F., Mehmood, M. H., Younis, B., Mehmood, N.,

Imran, T., and Zafar, U. (2022). Comparative analysis

of machine learning algorithms for classiﬁcation of

environmental sounds and fall detection. International

Journal of Innovations in Science & Technology.

https://journal.50sea.com/index.php/IJIST/article/vie

w/188.

Hawi, S., Alhozami, J., AlQahtani, R., AlSafran, D.,

Alqarni, M., and Sahmarany, L. E. (2022). Automatic

parkinson’s disease detection based on the combina-

tion of long-term acoustic features and mel frequency

cepstral coefﬁcients (mfcc). Biomedical Signal Pro-

cessing and Control, 78:104013.

Kumar, D., Satija, U., and Kumar, P. (2022). Automated

classiﬁcation of pathological speech signals. In IEEE

19th India Council International Conference (INDI-

CON), pages 1–5.

Feature Selection Improves Speech Based Parkinson’s Disease Detection Performance

731

Lamba, R., Gulati, T., and Jain, A. e. a. (2023). A speech-

based hybrid decision support system for early detec-

tion of parkinson’s disease. Arab Journal of Science

and Engineering, 48:2247–2260.

Majda-Zdancewicz, E., Potulska-Chromik, A., No-

jszewska, M., and Kostera-Pruszczyk, A. (2022).

Parametric analysis of pilot voice signals in parkin-

son’s disease diagnostics. Journal of Automation,

Electronics and Electrical Engineering, 4(1):21–28.

Polat, K. and Nour, M. (2020). Parkinson disease classiﬁ-

cation using one against all based data sampling with

the acoustic features from the speech signals. Medical

Hypotheses, 140:109678.

Priya, B., Rithiga, B., and R, B. (2021). Parkinson dis-

ease detection based on speech using various machine

learning models and deep learning models. In Interna-

tional Conference on System, Computation, Automa-

tion and Networking (ICSCAN), pages 1–6.

Priya, E., S, J. P., Reshma, P. S., and S, S. (2022). Tem-

poral and spectral features based gender recognition

from audio signals. In 2022 International Conference

on Communication, Computing and Internet of Things

(IC3IoT), pages 1–5.

Radovic, M., Ghalwash, M., Filipovic, N., and Obradovic,

Z. (2017). Minimum redundancy maximum relevance

feature selection approach for temporal gene expres-

sion data. BMC Bioinformatics, 18(1):9.

Sakar, B. E., Isenkul, M. E., Sakar, C. O., Sertbas, A., Gur-

gen, F., Delil, S., Apaydin, H., and Kursun, O. (2013).

Collection and analysis of a parkinson speech dataset

with multiple types of sound recordings. IEEE Jour-

nal of Biomedical and Health Informatics, 17:828–

834.

Tolosa, E., Garrido, A., Scholz, S. W., and Poewe, W.

(2021). Challenges in the diagnosis of parkinson’s dis-

ease. The Lancet Neurology, 20:385–397.

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

732