Abstract
Early detection of Parkinson’s disease (PD) through speech analysis offers significant clinical advantages, yet no validated tools exist for Arabic-speaking populations, representing a critical gap in global healthcare. Previous studies have relied on limited machine learning (ML) classifiers and voice attributes, which may introduce bias and hinder effective technique discovery. To address this, we developed an optimal PD prediction pipeline by testing multiple ML classifiers and feature extraction methods. We created the first Arabic PD speech dataset, comprising 40 subjects (17 with PD and 23 controls), and validated our methodology on an independent Spanish cohort of 100 subjects. Feature extraction included traditional, audio-to-text, and deep voice features from a pre-trained Whisper model. We employed feature selection and dimensionality reduction techniques to refine the dataset dimensions. Final features were assessed using twelve classifiers with leave-one-out and k-fold cross-validation for robust performance evaluation. Shapley additive explanations (SHAP) were utilized to determine feature importance as vocal biomarkers. Linear Discriminant Analysis achieved optimal performance with 90% accuracy, precision, recall, and F1-score using leave-one-out cross-validation. Linear Support Vector Classification also performed well, achieving 87.7% precision and 87.5% recall. When tested on the independent Spanish dataset, our methodology attained 83% accuracy, confirming cross-linguistic generalizability. SHAP analysis indicated that audio-to-text features provide contextual insights on fluency and coherence, while traditional features effectively capture acoustic variations. This study establishes the first validated Arabic PD speech classification system and demonstrates its universal applicability, laying the groundwork for global speech-based PD screening.
Plain Language Summary
Parkinson’s disease (PD) affects how a person speaks, often causing changes in voice quality and fluency that are hard for the human ear to measure in early stages. While technology can analyze these changes to help doctors diagnose the disease early, most of this research has been done in English or European languages. This has created a gap in care for over 400 million Arabic speakers worldwide. In this study, we developed the first validated system specifically designed to detect Parkinson’s disease from Arabic speech. We recorded speech samples from individuals with Parkinson’s and healthy volunteers. We then used advanced computer programs (machine learning) to identify patterns in their voices. To make sure our tool works for everyone and not just one group, we also tested it on a different group of people speaking Spanish. Our system was able to identify Parkinson’s with 90% accuracy in the Arabic group and 83% accuracy in the Spanish group. This research is important because it shows that speech-based tools can work across different languages and cultures. Because this tool can run on simple technology like a smartphone, it could make Parkinson’s screening much more accessible and affordable, especially in areas where specialized doctors are hard to find. This is a major step toward using technology to provide fairer healthcare for people with neurological conditions globally.
Keywords
Introduction
Dysarthria is a medical condition characterized by impaired voice production due to muscle control loss that leads to vocal effort or quality irregularities. 1 These issues significantly affect a person’s communication ability and reduce their quality of life. Symptoms of dysarthria include variations in the range, tone, strength, speed, duration, or precision required for speech control 2 The hallmark of the dysarthria is a reduction in speech intelligibility. 3 Uneven muscle tone and aberrant vocal fold oscillations are part of the pathogenesis in several cases. These conditions can be caused by hypertonicity, alterations in vocal fold mass, and insufficient glottis closure during phonation. 4 Within the framework of therapeutic practice, dysarthria is often considered a normal aspect of aging, 5 but it can also be a sign of a significant underlying disease, especially when individuals have dysarthria for more than four weeks and when it is linked to risk factors or another medical condition.
Hypokinetic dysarthria is a prevalent clinical type of dysarthria among its numerous variants. This is a result of the loss of dopaminergic neurons in PD. 6 Speech is often quiet, monotonous, and poorly articulated. One may observe a combination of hurried speaking and delayed speech onset. The afflicted individuals may also show additional distinguishing symptoms, including festinating gait, cogwheeling, masking of facial characteristics, and resting tremors. 7 The most common early indicator of PD is dysarthria. 8 Years before PD shows symptoms, a patient may develop dysarthria, mainly due to bending of the vocal folds or hypo adduction Consequently, measuring and detecting hypokinetic dysarthria is a valid and noninvasive approach to the early diagnosis of PD.
The global burden of PD varies significantly across linguistic and cultural populations, yet current speech-based diagnostic tools are predominantly validated in English-speaking cohorts. This linguistic bias creates healthcare disparities, particularly affecting Arabic-speaking populations numbering over 400 million worldwide. The development of culturally and linguistically appropriate diagnostic tools represents a critical unmet clinical need. 9
Currently, speech impairments are evaluated using a combination of neurological examinations and specialized assessments conducted by speech-language pathologists. Patients may undergo tasks that evaluate various elements of their speech, such as volume, prosody, and clarity. In the context of Parkinson’s disease, the Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) serves as a standardized, rater-dependent clinical tool that evaluates the general severity of symptoms, including speech impairments, among other motor and nonmotor symptoms. 10 Several studies indicated potential correlations between voice-based features and traditional UPDRS assessments.11,12 This method is effective but has certain limitations associated with the rater’s expertise and experience and the ”narrowness” of the scale (scores ranging from 0 to 4). This narrow scale can make it difficult to discern true differences between two patients who receive the same score, and the subjective nature of clinical observation may lead to a biased assessment.
There has been a paradigm change in recent years toward the use of spectral assessment-based tools for the objective examination of voice samples (e.g., Cepstral/Spectral Index of dysarthria). The advantages of cepstral measurements in the evaluation of dysarthria have been reported 13 On the other hand, the human voice is a complicated trait with complex components. Therefore, in addition to the independent use of cepstral analysis of particular voice features (i.e., fundamental frequency), more sophisticated methods that are capable of analyzing and dynamically combining high-dimensional datasets of voice features, such as artificial intelligence approaches, may greatly increase the accuracy of the objective classification of voice samples in PD. The strength of ML approaches lies in their ability to identify latent patterns in physiological signals, which can help clinicians make accurate diagnoses of dysarthria and connect it to PD. In this context, ML algorithms can be trained on a voice dataset to create models that can automatically interpret new data and forecast future results, thus helping to identify patterns and trends that could be challenging to determine clinically. Idrisoglu et al. 14 critically evaluated several studies that discuss the validity of ML approaches for evaluating the auditory and perceptual characteristics of audio recordings. As their use lowers inter-observer variability and increases the precision of classifying PD speech samples, ML approaches are highly valued and offer additional advantages for the diagnosis of PD.
Many factors, such as subject-specific traits or ambient circumstances, may influence the human voice. The impact of voice features on speech output, particularly in the context of PD, remains largely unexplored and has often been described qualitatively. The majority of analyses have focused on vowel phonation, which was spoken for a few seconds straight during the acoustic examination of the signal. 15 The associated fundamental frequency (F0), signal amplitude, and resonance frequencies in the vocal tract or formants (Fn) all contribute to its quasiperiodic audio signal. Little et al and Tasnan et al.16,17 have offered a toolset that categorizes the traditional characteristics and metrics used to assess the most common issues related to the dysfunctional voices of PD patients. These features are extracted directly from the voice signal, including jitter, shimmer, HNR, and MFCC. Most patients on whom this procedure is used and validated speak English. Nevertheless, vocal characteristics may vary according to the language spoken. 18 It is valuable to provide a clear comprehension of the speech content and, of course, for many languages, some other important characteristics that are missing from those reflected in the existing tools could be further studied.
Contextual information from continuous speech would be helpful in the diagnosis of voice issues via vocal-text transcription. 19 Instead of diagnosing how PD affects the acoustic features of speech, the strategy is to take advantage of the ability to transcribe audio files by feeding them into an ML model, presuming that a person suffering from dysarthria is expected to speak less accurately, which causes the model to malfunction when transcribing their audio signal. The distinction of the PD would, thus, be aided by determining any variations in the pattern of effects among instances.
The use of ML for diagnosing dysarthria in PD is not uncommon. 20 Despite its proven effectiveness, variable performance between different types of classification methods has been reported.21,22 The reason for this is variations in the data and the model-driven nature of the various methods. Moreover, previous studies used a small number of ML classifiers depending on a limited set of voice attributes, which may induce bias and/or limit the search for more efficient ML techniques. Utilizing low-interoperability models may prompt doctors to express reservations when required to provide high-level evidence in clinical practice. 23 A lack of universality and the over-fitting problem could result in such a situation. This means that testing a variety of ML classifiers and feature extraction methods is needed to produce an ideal landscape to select the system that best fits the prediction of PD.
Moreover, it is crucial to confirm whether feature extraction techniques and the number of used features influence the performance of the considered ML models. Relevant studies have employed feature selection techniques to choose relevant variables with corresponding significance measures and improve prediction accuracy. 24 Random Forest (RF) is one of the most commonly used algorithms for this purpose. However, no research has yet examined the impact of feature selection versus integrating every feature on the precision of ML-based PD evaluation. Furthermore, cross-validation—the de facto norm in traditional ML studies—and dimensionality reduction of the data have only been carried out in a limited number of experiments, and most of the research was limited to hold-out validation. Feature selection, when combined with feature reduction techniques, can effectively handle high-dimensional data. This is helpful as many features are frequently present in datasets in the medical field, some of which may be superfluous or unnecessary. Direct analysis of such data may result in overfitting, increased model complexity, and processing inefficiencies.
Although ML has enormous potential to increase the effectiveness and caliber of speech data interpretation, significant obstacles still exist. Among these, the interpretation of the plot for pathologists and the decision-making process are arguably the most crucial. The application of SHapley Additive exPlanation (SHAP) to get around this problem is a significant advancement in methods for interpreting and explaining ML models. 25 This cutting-edge Python module is frequently utilized in ML applications during the feature engineering phase. To explain the outcomes of ML models, it links optimal credit allocation to local interpretation using the traditional Shapley value of game theory and its extension.
The aforementioned gaps inspired us to conduct a thorough study of the existing literature, with an emphasis on feature extraction methods. The majority of published papers have used very modest feature sets, with few exceptions, which were primarily drawn from small toolkits and pertinent approaches. Furthermore, the application of Mel Frequency Cepstral Coefficients (MFCC) has been employed.21,26 These feature sets were chosen for their ability to effectively characterize dysphonic speech and their reputation for capturing perceptual information. In their study, Alalayah et al 27 examined voice signals, developed by Oxford University’s Max Little, with respect to the UCI sustained vowels, comprising speech samples from both PD and Healthy Controls (HC). They evaluated the database using ML classifiers and extracted the 23 auditory features described by Little tools kit. Giuliano et al presented an ML technique for PD diagnosis using 339 acoustic features utilizing the vowel “a”. 26 They employed the dysphonia feature metrics suggested by Tsanas et al.. 17 Costantini et al. 21 suggested the use of a variety of traditional features in addition to non-linear features, including pitch period entropy and glottal-to-noise excitation, vocal parameters pertaining to the inadvertent low-frequency vibration of the vocal fold, vocal formats, and their energy. They assessed end-to-end classifier topologies and the performance of both conventional and DL pipelines in order to categorize PD and HC using 439 features. In a similar vein, Yuan et al. 28 prepared a custom dataset obtained from a PD cohort, consisting of 8 main features and 747 feature variables, including baseline features, time-frequency features, MFCCs, wavelet transform-based features, and vocal fold.
ML classifiers and deep neural networks have been used to classify PD and HC. Suppa et al. 29 investigated the use of support vector machines (SVM) to assess the impact of disease severity on PD voices at the early and mid–advanced stages under ON/OF medication status. Features were gleaned from the INTERSPEECH2016 database in order to assess the voice tremor of PD patients. On the other hand. Alshammri et al. 20 constructed ML models using the Synthetic Minority Over-sampling Technique (SMOTE), Feature Selection, and hyperparameter tuning (GridSearchCV) techniques, utilizing the voices dataset from the University of Oxford repository established by Little et al.. 16 Karan et al. 30 suggested using Hilbert cepstral coefficient (HCC)-based features in conjunction with a combined VMD and HS technique. Warule et al. 31 examined the discriminating capability of the time-frequency analysis of speech signals to categorize PD and HC individuals using sustained vowels and words. Moreover, a comprehensive systematic review conducted by Gelderen et al., 32 synthesized and discussed all recent speech-based deep learning methods for PD classification, contributing valuable insights into the advancements in this field.
Recently, transformer-based architectures have gained considerable traction in speech-based PD detection. La Quatra et al. 33 proposed a bilingual dual-head deep model for PD detection from speech, demonstrating that transformer features can capture cross-linguistic speech biomarkers. Similarly, Hemmerling et al. 34 explored multilingual PD classification through voice analysis using transformer-based representations, achieving improved interpretability for clinical applications. While these end-to-end deep learning approaches have shown strong results, they typically require large training datasets and substantial computational resources, 32 which may limit their applicability in clinical settings with small patient cohorts. In contrast, the present study adopts a more interpretable ML pipeline that leverages pre-trained model features (Whisper) combined with traditional acoustic features and classical ML classifiers, offering a practical and transparent alternative that is well-suited for smaller clinical datasets.
A limited number of ML approaches were considered in the pertinent investigations, which could lead to bias and/or restrict the search for more effective ML methods. This motivates us to conduct this study using 12 well-known classifiers with independent classification schemas and contrast their performance on voice data. These are Neural Networks, Linear-based approaches, and Decision Tree. By assessing the performance of each classifier using thorough Leave-one-out cross-validation (LOO-CV); in addition to k-fold cross validation to provide predictions, creating the ideal setting for choosing the model that best matched the given data. Additionally, we empowered our study through the use of three sets of features. The first set included the most relevant traditional features directly extracted from the speech signal, including the well-known MFCC coefficients. The second and third sets of features were extracted from the audio signal utilizing a supervised Pre-trained Speech Recognition model; namely, the Whisper model. 35 All these features were merged and fed to each ML method for the classification of the degree of dysarthria.
Recent research reveals that PD-related speech patterns may cross linguistic barriers, implying the possibility of universal diagnostic indicators. However, this hypothesis requires rigorous cross-linguistic validation to ensure clinical applicability across diverse populations. 36
Lastly, by utilizing the SHAP package to examine the universe of features, this work aims to challenge the ”black box” model that guides the use of autonomous ML approaches. Consequently, being aware of how each feature contributes to the models’ ultimate classification of PD and healthy voice in suspected patients. We hypothesized that, through the use of the proposed pipeline, we would be able to investigate the best ML approach for dysarthria assessment in PD patients.
Methods
This research was conducted and reported in accordance with the STROBE (STrengthening the Reporting of OBservational studies in Epidemiology) guidelines, which is normally used for cohort, case-control, and cross-sectional observational studies (https://www.strobe-statement.org), and with reference to the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis) statement for prediction model studies. We have ensured that all relevant reporting elements, such as study design, participant eligibility criteria, sample size, feature extraction procedures, model evaluation, and study limitations, are addressed in the respective sections below.
Patient Cohort
Patients were recruited in collaboration with the Parkinson Disease Patients Society, from the neurological departments of Al-Karak Governmental Hospital and King Abdallah Hospital. The three entities issued an invitation letter to the potential participants based on their registries. More than 800 potential registered patients from the society and clinics are estimated to have been invited to participate in the study.
Seventeen PD patients and 23 healthy subjects (HC) were included in the control group. While we recognize that a higher sample size is preferable, we decided that 23 normal instances are adequate to maintain a balanced dataset. The registry’s old status accounts for the poor response rate among PD patients. For example, we discovered missing addresses, and several patients were either residing outside Jordan or had passed away. Furthermore, the majority of patients were hesitant to join, notably their guardians, namely their sons and daughters. Privacy concerns frequently motivate this hesitation, as they do not want a stranger to hear about their parents’ experiences or sensitive information.
The Jordanian patient cohort consisted of 17 individuals (10 males and 7 females) with varying stages and symptoms of Parkinson’s disease, ranging in age from 43 to 77 years old. All were on medications for Parkinson’s during the period when our data was collected from them. The control group comprised 23 healthy individuals (13 males and 10 females), aged 38 to 80 years old. The cases were typical of Jordan’s northern, southern, and central regions. All participants (control and PD patients) were native Arabic speakers and non-smokers. None of the participants reported bilateral/unilateral hearing loss, respiratory disorders, or other non-neurologic disorders affecting the vocal cords.
Parkinson’s Disease Speech Datasets Comparison
As it was difficult to record audio for more PD patients, we initially attempted to increase the number of PD records by including recordings from 4 Moroccan dysarthria patients. This was done to increase the number of positive examples, with Moroccan patients uttered almost the same text as the Jordanian PD patients. However, we later removed these additional four records after observing a significant increase in classification accuracy. Because the system appeared to learn the accent differences rather than the characteristics of PD in the voice. This was particularly evident when considering the large differences between the Moroccan and Jordanian dialects, which can vary significantly in pronunciation, intonation, and rhythm. 42 Such variations may surpass the subtle vocal changes associated with PD, leading to a potential misinterpretation of the data. As a result, the system’s ability to accurately identify PD may be compromised, as it focuses more on regional accents rather than the clinical markers of the disease.
Recording audio from more PD patients in Jordan proved difficult because of a number of practical, social, and cultural issues. Because the PD is stigmatized in society, many families were reluctant to let their loved ones participate in these recordings. As Parkinson’s disease may impair speech and movement, several families were afraid that disclosing their loved one’s illness would cause shame or humiliation in their community. Furthermore, PD patients’ guardians frequently voiced worries about the patients’ mental health. Given the unpredictable nature of the illness, they were concerned that their loved ones would experience stress or discomfort throughout the recording process. The belief that taking part in research may raise awareness of their condition—which some patients want to keep private—exacerbated this worry. Another major worry was privacy. Families were concerned that the recordings would be misused and that private information might be unintentionally revealed. Sharing such personal information about a family member’s illness might be viewed as a breach of trust in a society where mental and physical health conditions are frequently stigmatized. The matter was further compounded by stringent confidentiality standards among healthcare practitioners. Many physicians were hesitant to include their patients in studies that would reveal private information. A great degree of trust is usually present in the doctor-patient relationship, thus any perceived invasion of privacy may deter patients from getting treatment or taking part in further research. Lastly, despite our best efforts throughout the data collecting phase, patients and their families were not aware of the significance of taking part in research investigations. Many people were not entirely aware of how their participation may improve our knowledge of and ability to treat Parkinson’s disease. Promoting involvement was made much more challenging by this unfamiliarity.
The difficulties connected with gathering speech datasets for PD are well established, as indicated by the scarcity of datasets accessible in the literature. Most known datasets range in size from 31 to 100 subjects, showing the challenge of collecting enough data in this discipline. Our 40-subject dataset fits well within these recognized standards and contributes significantly to Parkinson’s disease research. Notably, all datasets papers included in our investigation have been validated to include new data rather than simply evaluating pre-existing datasets. We confirmed the subject’s numbers and languages from the original sources and cross-checked all references with primary publications. In this context, our dataset of 40 participants is comparable to the Sakar et al dataset, 40 which also has 40 subjects, and it outperforms the MDVR-KCL dataset, 41 which has 37 subjects. Furthermore, our dataset is the first Arabic PD speech dataset suitable for classification tasks, which fills a major gap in multilingual research by addressing the variability of PD speech patterns across languages. Thus, our dataset’s sample size is not only justified by field standards, but it also provides a valuable and adequately scaled addition to existing research in PD speech analysis, as shown in Table 1.
While bigger datasets are preferred, our 40-subject cohort is consistent with known PD speech research criteria, outperforming the Sakar et al dataset (40 individuals) and the MDVR-KCL dataset (37 subjects). Critically, our validation of Little et al's features yielded 65% accuracy, which is significantly higher than chance (50%), confirming meaningful PD-related signal detection in our dataset.
List of the Audio Dataset of PD Patients and the (Normal or Healthy) Control Group (HC)
Because we are dealing with a classification problem based purely on voice recordings, both groups must be as sex- and age-matched as feasible. Otherwise, the proposed system may learn to differentiate between groups based on sex-related vocal features rather than recognizing PD vs HC. Similarly, age can impact vocal features, since younger participants tend to have unique voices compared to older persons. As a result, we must examine and execute statistical tests on these groups to ensure that there is no age and sex bias.43-45
By performing a t-test on the age data of both groups (PD vs. N), the results show that there are no significant differences between the PD group and the HC group, as the age t-test yielded a t-statistic of 1.1146 and a p-value of 0.2730, indicating that the age distribution between the two groups does not differ significantly, indicating any bias due to age.
Similarly, the Chi-squared test for sex distribution revealed no significant difference. The contingency table showed that the HC group had 10 females and 13 men, whereas the PD group had 7 females and 10 males. The Chi-Square statistic was estimated as 0.0, with a p-value of 1.0000. This investigation demonstrates that there is no substantial difference in sex distribution between the two groups. Table 2 and Figure 1 depict some statistical characteristics that further support the demographic unbiasedness of our sample. Distribution of age and sex across the two groups PD and HC or normal (N)
Statistical Characteristics of the Demographic Information of the Participants in Both Groups, PD and HC
Voice Recordings
Voice recordings were collected by asking participants to carry out a specific speech task with their usual voice intensity, pitch, and quality. The speech tasks consisted of the sustained emission of the vowels of the Arabic alphabet and some of the most frequent words and sentences, including. • (”alif” with ”fathah”): a • (”alif” with ”dammah”): u • (”alif” with ”kasrah”): i • (”alif” with ”maddah”): ā • (Triple “waw”): oo • (triple “yaa”): ee • Arabic numbers from 1 to 10 • Surah ”Al-Fatiha”: This is the first chapter of the holy Quran, which is considered as one of the most frequently recited and memorized chapters by Muslims, as the targeted people were all Muslims.
We used these words to be recorded because they could be easily memorized by all participants, do not require reading capabilities, and have been utilized in existing literature for the development of Arabic voice pathology dataset.
46
Importantly, the standardized nature of these texts ensures consistency across recordings, minimizing the influence of regional accents. Unlike Moroccan accents, the various Jordanian accents—whether from the northern, southern, or central regions—do not significantly affect the analysis in terms of accent variation. This is because the text being uttered is uniform, allowing for consistency in pronunciation. For instance, when reciting a verse from the Holy Quran, all Jordanians pronounce it almost the same in Modern Standard Arabic, similarly, they pronounce the numbers (1-10) in the same way, ensuring that our dataset reflects a true representation of PD speech characteristics without the confounding effects of accent diversity.47,48 Figure 2 shows the text spoken by all participants in Arabic font. The text spoken by the participants, in arabic font
Voice recordings were collected using a Sony ICD-TX50 audio recorder, a digital voice recorder designed for high-quality recording in a small and elegant package. It has noise reduction, excellent sensitivity for audio recording, and simple computer connectivity for file sharing. Due to its advantages, it was considered a good choice for documenting lectures, meetings, interviews, and personal notes. 49
Regarding data preprocessing, we did not apply audio filtering, noise suppression, or signal enhancement to the recorded speech signals before the feature extraction. This was a deliberate design choice intended to preserve the natural acoustic characteristics of the voice, including those subtle irregularities and distortions that are hallmarks of dysarthric speech in PD patients. Applying audio filters could inadvertently smooth out or remove the very vocal artifacts, such as tremor, breathiness, and reduced loudness, which might be vital for distinguishing PD from healthy speech. Thereby diminishing the discriminative information available to the classifiers. The raw audio signals were used directly as input to the feature extraction pipeline, ensuring that the extracted features reflect the authentic vocal characteristics of each participant.
The Proposed ML System
Our goal in this work is to devise an intelligent system to identify Parkinson’s patients from their audio signals. To achieve this goal, the proposed method is designed to classify whether a patient has PD or not based on the degree of dysphonia determined in their speech signal.
The methodology of the proposed model is structured into four steps: feature extraction, feature selection and dimensionality reduction, machine learning classification models, and performance evaluation. In addition to several machine learning Python libraries such as scikit-learn, TensorFlow, and Keras, Weka (Waikato Environment for Knowledge Analysis) was used to perform the analysis.50-52 The steps of the proposed framework are shown in Figure 3. The flow diagram of the proposed Parkinson’s diagnosis system based on dysphonia
Feature Extraction
In order to achieve the goal of this study, we aimed to determine the best set of features and the optimal machine learning method that can be learned from our data to achieve the best possible prediction performance. To this end, we investigated three sets of features. 1. Traditional features that are directly extracted from the voice signal (TF set); 2. Features extracted from the text result of a pre-trained Audio-to-text deep learning model (A2T set); 3. Deep features extracted from the last layer of a pre-trained Audio-to-text deep learning model (DEEP set).
TF Set
These features were extracted directly from the voice signal. Here, we opted to extract 40 features of the Mel Frequency Cepstral Coefficients (MFCC). The coefficients that add up to an MFC are called MFCCs, which originate from a particular kind of Cepstral representation (a nonlinear “spectrum-of-a-spectrum”) of the audio sample. MFCCs are commonly used in voice recognition and various other audio analysis processes.53-56
It is recommended that a maximum of 25 MFCCs are used for voice analysis, even though many of the studies that have already been conducted have only employed 13. 57 Additionally, it has been confirmed that improved classification metrics are obtained when increasing the number of MFCCs. For this reason, we opted to use 40 coefficients, knowing that we were going to perform further feature engineering to decide which of these features were to be used in the final system.
Dysphonia may involve a variety of factors including rhythm, pace, pronunciation accuracy, and the presence of hesitations or fillers, which might not be fully captured by MFCCs alone.
Therefore, we added more features to this set, including. 1. The mean of the pitch of the audio signal and its standard deviation. Pitch is a perceptual feature that enables the arrangement of sounds on a frequency-related scale that runs from low to high. 2. The energy of the audio signal, which represents the loudness or intensity of the sound and can be calculated using Equation 1 3. The mean of the zero crossing rates (zcr), which calculates the rate at which the signal changes from positive to negative or back, may provide insights into Dysphonia as a speaker with Parkinson’s may have longer pauses and more frequent hesitations. 4. The Articulation Rate, which reflects the number of syllables spoken per second, and can be approximated using Equation 2 5. Number of pauses and silence duration. For simplicity, we detect silent intervals using a threshold-based approach to energy. 6. Speech variation, which can capture the rhythm present in an audio signal.
These features comprise the first set of features including 40 MFCCs and the 8 additional features detailed above, which accumulate to 48 features to be used as input to a machine learning method. These features were extracted from the audio signals using Librosa, a Python package for audio analysis. Algorithm 1 shows the pseudo-code for extraction in the first set of experiments.
A2T Set
When attempting to distinguish dysphonia in Arabic speakers, leveraging a pre-trained deep learning model specialized in speech processing could significantly improve the results. These models have been trained on vast amounts of data and can capture complex speech patterns, intonations, and nuances that may characterize dysphonia.
In this work, we opted for OpenAI’s Whisper, due to its availability and its state-of-the-art performance. This model can transcribe speech from audio into text across a wide range of languages, including Arabic. The idea is to feed the audio files into the Whisper model for transcription and then to analyze the resultant text for Dysphonia. Assuming that a person with dysphonia is likely to be a less accurate speaker, Whisper may fail in transcribing their audio signal accurately.
We decided to use the Whisper model for extracting additional features from audio recordings due to its resilience and adaptability in handling various audio sources. This pre-trained speech recognition algorithm is intended to transcribe spoken language under a variety of settings, which is critical for assessing speech patterns in people with dysphonia. 48 As a result, it is vital to highlight that persons with dysphonia may speak less accurately, which may have an influence on transcription quality and hence the extracted features, allowing distinguishing features between HC and PD participants. While there are other models available, Whisper’s effectiveness in noisy conditions and flexibility to multiple languages make it a good fit for our investigation. 49
The output of the Whisper model, given an audio signal, is a dictionary data structure containing the following. 1. Text: The transcribed text itself, which can be analyzed further to provide direct insight into the speech content, as more accurate text should be associated with Normal speakers, while less accurate text may be associated with Dysphonia. 2. Segments: Each segment contains detailed information about specific portions of the audio, including: • The start time of each segment (seconds). • The end time of each segment (seconds). • The segmented Text. • The Tokens (numeric representation of each word). • The average log probability of each predicted token in the segment. This indicates how confident the model is in its prediction for each word. • Compression ratio, which denotes the compression achieved by the Whisper model compared to the original audio size. • No-speech probability, which indicates the likelihood that the segment consists of silence.
Figure 4 shows a Whisper sample result of a HC speaker. A whisper transcript sample result of a dysphonic subject showing four segments, we covered some text by yellow to avoid showing the holy qur’anic verse that is wrongly written
As can be seen from Figure 4, the transcribed text of a random dysphonic speaker exhibits several mispronounced words—for example, the Arabic words for one, four, eight, and ten. Such errors indicate a certain level of dysphonia; a lesser number of errors may indicate that the subject is more likely a normal speaker. However, this condition can be violated even for normal speakers, due to many reasons such as noise, speed of taking, accents, and so on; nevertheless, dysphonic speakers should be more likely to have less accurate transcriptions than normal speakers. Therefore, we used the accuracy of the Resultant Text to extract the following features. 1. Word Error Rate (WER): This error is determined by comparing the resultant text to the original spoken text using the JiWER library, which is a simple and fast Python package designed to evaluate an automatic speech recognition system. 2. Character Error Rate (CER): Similar to the WER, but calculated at the character level. 3. Levenshtein Distance (LD): The number of edits needed to make the resultant text equal to the reference spoken text. 4. Simple match percentage (SMP): This calculates the match percentage between the resultant text and the reference spoken text in terms of words. 100% means no error, and 0% indicates that all of the transcribed text is wrong for a particular speaker. 5. The number of words: A simple count of the number of words found in the resultant Text.
From the resultant segments, we extracted the word count, the average log probability, Compression ratio, and No-speech probability features provided by the segments; however, as each speaker had a different number of segments, ranging from 1 to 20 segments, we opted for statistical description of each feature, using the summation, minimum, mean, median, and maximum of each feature. Instead of extracting the start and end times, we calculated the duration of each segment, which is equal to (end time − start time) and statistically described it similarly. As dysphonic speakers generally speak slower than normal speakers, we also calculated the speed of speech, which indicates the number of words spoken per minute, as follows:
The number of features extracted from the resultant dictionary of the Whisper model after inputting each audio file was as follows: • Six features extracted from the recognized Text; namely, WER, CER, LD, SMP, the number of words recognized, and speech speed. • Five statistics (summation, minimum, mean, median, and maximum) × 5 features (i.e.,number of words in each segment, the log probability of each segment, Compression ratio, No-speech probability, and the duration of each segment) = 25 features.
However, we excluded the summation of the number of words found in each segment, as the number of words was already extracted from the text, and we excluded the summation of the duration as such information is embedded in the speed equation. Therefore, the total number of features in A2Tset was 29.
DEEP Set
Deep learning has been widely applied to extract features, especially extracting deep features from pre-trained deep models such as those mentioned in 58-62. Although they are not interpretable, deep features could be useful for describing different aspects of an audio file. These features can typically be extracted from any layer (except the input layer) of a pre-trained model. Here, we opted to use layer 5 of the Whisper model and recorded the 512 values there as features for each speaker, without fine-tuning the model. All the previous sets of features are combined into one file, comprising 48 + 29 + 512 = 589 features of all types.
Feature Selection
We employed the widely used RF feature importance for the feature selection investigation, to determine the best variables for use in the machine learning context.63-65 Figure 5 shows the features, ranked according to their importance. The top 57 features, sorted in descending order based on their importance for machine learning using RF. Total importance accumulates to 60% of importance
To establish a threshold for selecting the most important features, we identified those that collectively accounted for at least 60% of the total importance. This resulted in the selection of 57 features based on their importance scores. Among these, 4 were TF features (MFCC4, MFCC8, Pitch mean, and zcr mean), and 6 A2T important features (Mean & median average log probabilities, CER, summation of no speech probabilities, Levenshtein distance, and Maximum average log probabilities). The remaining were DEEP set features, as illustrated in Figure 5.
It is worth noting that, due to the stochastic nature of the Random Forest method, some features appeared in the top 60% in different runs. These features, which may vary with different random seeds, include (but are not limited to) the number of pauses and mean duration, this indicates the importance of all sets of features.
It is worth noting that the best 57 features, as shown in Figure 5, came from all three of the proposed feature extraction approaches, which justifies merging these features to form one dataset for training the prediction system. Therefore, the top 57 features were used for the subsequent prediction analysis, and the outcomes were compared against predictions made using the whole dataset and other dimensionality reduction methods.
Dimensionality Reduction
With so many dimensions (589 features), a machine learning method is required to request additional examples for better learning. Since we only have 40 examples, we have to restrict the amount of features to enhance learning from such a small sample.
Dimensionality reduction has three main benefits: it makes machine-learned models smaller, requires less computing power during the training and testing stages, and allows for training from small samples as the feature space becomes smaller. Therefore, we utilized Principal Component Analysis (PCA),66,67 Spectral Embedding (SEM),59,68 and t-distributed Stochastic Neighbor Embedding (t-SNE),60,69 three widely used techniques for dimensionality reduction. Typically, these methods play a crucial role in reducing a dataset’s dimensionality while maintaining its fundamental structure, which helps with further analysis and model construction.
PCA is a linear dimensionality reduction approach used for data pre-processing and visualization. The data are transformed into a new coordinate system, allowing the directions (principal components) that capture the most variation in the data to be identified. 58 PCA consists of three primary phases: covariance matrix calculation, eigenvalue and eigenvector extraction, and principal component selection.
As can be seen in Figure 6, the first 7 components preserved at least 99% of the data variance; however, we retained the top 10 principal components in PCA, which together preserved 99.49% of the variance in the data, in order to keep more data for the learning process. The top 7 principal components preserve at least 99% of data variance
For a dataset D, the covariance matrix C is calculated as follows:
Then, the eigenvectors (V) and eigenvalues (λ) can be calculated from Equation 4 as follows
After that, the principal components (PCs) are formed by choosing the eigenvectors corresponding to the top k eigenvalues
One important point to mention here is that PCA is applied only to the training set, not to the entire dataset (train and test) to avoid data leakage. In the case of cross-validation, PCA is computed on the training set for each fold, and the test set is then transformed using the eigenvectors derived from the training set, preventing data leakage and ensuring proper separation between training and testing data across all folds.
Our data were represented in a much lower-dimensional space using the Principal Components (PCs) obtained from this method. The complexity and computation time of machine learning models can be greatly reduced through such a reduction in dimensionality. More significantly, this approach seeks to improve the model’s capacity to identify patterns with the aim of providing better predictions.
The first two components of the aforementioned transforms are often the most important dimensions of the transformed dataset. Figure 7 illustrates the distribution of these two important components obtained by the four transforms across both classes (Normal = 0 and Dysphonia = 1). Data visualization using the top two components of each transform
Figure 7 demonstrates a low degree of class overlap, suggesting a high degree of separability regardless of the type of transformation applied (linear or non-linear); however, it should be noted that this may be due to the small number of examples, and not the presence of clear patterns to be identified in the data.
Experiments Setup
To determine which ML classifier best suits our data, the first set of experiments was conducted by examining 12 classifiers on all of the extracted voice features, including SVM, K-nearest neighbors (KNN), Hassanat KNN (HKNN).70-73 Linear discriminant analysis (LDA), Extreme Gradient Boosting (XGBoost), Linear SVM, Classification And Regression Tree (CART), RF, ANN, Naive Bayes, AdaBoost, and Stochastic Gradient Descent (SGD).
K-fold cross-validation (CV) has been frequently employed in scientific research for many years due to its benefits over hold-out set CV. It enables the use of the complete dataset for both training and testing sets, with the accuracy metrics averaged across the k runs. The 10-fold and 5-fold CV are popular options among researchers, therefore, we used both in some experiments; however, we mainly focused on the 40-fold CV, i.e. leave-one-out cross-validation, since it allows for a larger portion of the data to be trained and test all examples as well. Given the limited size of our dataset, a bigger training set provides for more rigorous learning and allows for better model performance on unseen data, which could improve the generalizability of the results.74-76
Accuracy Results for Different Machine Learning Models, Using 10-Fold Cross-Validation. The italic Font Indicates the Best Performance
Results
By analyzing the data in Table 4, it can be seen that the performance of most classifiers was good, with average accuracy ranging from 62.5% to 87.5%. In particular, LDA outperformed other classifiers for PD classification, with the greatest mean accuracy of 87.5%, a median accuracy of 100%, and the lowest standard deviation of 0.177. In comparison to the other method, the linear classifiers had the best performance. In particular, LDA showed outstanding consistency, with perfect accuracy in 50% of the runs and a narrow interquartile range (IQR) of 75% to 100%. Following LDA, LinearSVC, and AdaBoost performed well with mean accuracies of 82.5% and 80%, respectively, demonstrating great stability and high maximum accuracy. This finding is also evident from Figure 8, where these methods were more stable across the 10 runs when compared to the others. Box and whisker plots and density of accuracy for comparison of classifiers across the 10-fold cross-validation process
Mid-tier models including HKNN, SGD, RF, and ANN performed moderately, with mean accuracies ranging from 75% to 77.5%. Among them, RF had the least standard deviation (0.167), demonstrating reliability despite lower mean accuracy. In contrast, HKNN and SGD had larger IQRs, indicating sensitivity to data splits, but ANN had considerable variability (standard deviation of 0.264), which was most likely due to initialization and hyperparameters choices.
Lower-performing models included Support Vector Classifier (SVC) and XGBoost, with mean accuracies of 72.5% and 67.5%, respectively, while CART and KNN performed the worst, with high standard deviations and low median accuracies.
LOO-CV Results for PD Classification Using the Best Performers Across Different Dataset Scenarios, Highlighting Precision (P), Recall (R), and F1 Score (F). Best Performances are Indicated in italic Font
The top performers demonstrate varied levels of performance in diverse data contexts, as Table 5 illustrates. When using all of the extracted features, the LinearSVC classifier performs best, achieving 90% for F1 Score, Precision, and Recall. The LDA classifier, on the other hand, performs a bit lower with 87.5% for Precision, Recall, and F1 Score, indicating a preference for the selected 57 features.
Using either the whole feature set or the selected feature set produced from the proposed feature extraction methods consistently yields the best results, regardless of the classifier employed. This demonstrates how important the three proposed feature extraction techniques are for improving PD classification from speech recordings.
When used alone, the proposed feature sets offer high to moderate performance, indicating considerable potential. Notably, for Precision, Recall, and F1 Score, no feature set produced a performance lower than 0.566. With Precision, Recall, and F1 Score values of 0.836, 0.825, and 0.826, respectively, the Deep Features feature set (alone) outperformed the other feature sets. The A2T features come next, with Precision, Recall, and F1 Score values of 0.7745, 0.775, and 0.774, respectively. The precision, recall, and F1 scores of the TF features were 60% - 73%, indicating a reasonable level of performance. The robustness of the proposed feature extraction approaches is demonstrated by their richness in features that are significant in accurately identifying PD subjects from speech signals.
Notably, the dimensionality reduction techniques performed poorly, most likely as a result of losing necessary information. The PCA approach, which has shown success in maintaining data variance, is the exception to this rule. We were able to obtain a thorough depiction of the data by choosing the top seven principal components. The relevance of the proposed three feature sets is further highlighted by the fact that these primary components were obtained utilizing all of the features that were available.
Results Benchmarking
Arguably, the first study to determine whether dysphonia measurements are appropriate for Parkinson’s disease telemonitoring was conducted by Little et al.,
16
and approximately the same features have been employed in the majority of studies conducted since. We contrast our suggested approach with the features of Little et al.’s collection.
16
Each of these features contributes a single value per audio file, for a total of 10 extracted features for each file processed; namely. 1. Jitter (Absolute); 2. Jitter (DDP); 3. Shimmer (APQ); 4. Shimmer (DDA); 5. Noise-to-Harmonics Ratio (NHR); 6. Harmonics-to-Noise Ratio (HNR); 7. Recurrence Period Density Entropy (RPDE); 8. Detrended Fluctuation Analysis (DFA); 9. Correlation Dimension (D2); 10. Pitch Period Entropy (PPE).
We extracted these features, as described in Little et al.’s collection, and used the extracted features for comparisons using the top performer classifiers as shown in the last row of Table 5.
Table 5 demonstrates how the proposed features utilizing the three sets of features significantly -in all scenarios-outperformed Little et al.’s features in all metrics. This is also evidenced by the relatively low performance of these features as shown in Figure 9, if compared to using the proposed set of features as shown in Figure 10. The classification result obtained on our dataset utilizing the 10 Little’s features, which attained a performance rate (P, R, and F) of 60-65%, validate our small yet high-quality dataset. This result demonstrates that our dataset is more than just random noise; it successfully captures distinguishable voice differences associated with PD. The results indicate that our methodology is sound, with effective feature extraction and a robust machine learning pipeline. Importantly, this emphasizes the idea that quality is more important than the quantity when it comes to dataset efficacy.78-81 Confusion matrices and AUC of both of the top classifiers using LOO-CV on Little’s et al extracted features Confusion matrices and AUC of both of the top classifiers using LOO-CV on the proposed features (the whole set)
In our work, where the baseline random chance performance is 50% (PD or HC), while Little et al's features performed at 65% confirms that our dataset contains meaningful signals related to PD. Notably, These features were originally designed for English-language PD subjects, yet they surprisingly managed to modestly identify PD in Arabic speakers. Our methodology outperformed Little et al's features results with an accuracy of 90%, demonstrating that our innovative approach enhances detection capabilities. This three-tier evidence—comparing random chance, Little’s features, and our own—provides powerful validation for our dataset. The accuracy of 65% suggests that while detectable PD patterns exist, they can be challenging to identify, reinforcing our strategic position that a smaller, high-quality dataset can yield significant insights into PD voice characteristics.
To further quantify the reliability of our findings, we have computed 95% bootstrap confidence intervals using 10,000 iterations with stratified resampling. This is done mainly for the top-performing methods. For LDA using all 589 features, the accuracy of 90.0% yielded a 95% CI of [80.0%, 97.5%], with precision 0.900 [0.800, 0.976], recall 0.900 [0.800, 0.975], and F1 0.900 [0.798, 0.975]. For LinearSVC using the 57 selected features, the accuracy of 87.5% yielded a 95% CI of [77.5%, 97.5%], with precision 0.877 [0.777, 0.976], recall 0.875 [0.775, 0.975], and F1 0.875 [0.774, 0.975]. Such confidence intervals confirm that classification performance is substantially above chance (50%) even at the lower bound. Furthermore, McNemar’s test was applied to compare our proposed features against Little et al.’s features under LDA classification. The test revealed a statistically significant improvement (p = 0.007), confirming that the performance gain of our methodology (90.0% vs. 60.0%) could not be attributable to chance.
In order to further validate our methodology on a different PD dataset, we opt for the largest available PD dataset (PC-GITA) provided by Juan Rafael Orozco Arroyave and colleagues, who graciously granted us access to their recordings. 38 We applied our methodology to PC-GITA dataset comprising speech recordings from 50 individuals with PD and 50 healthy controls. The participants recited the same dialogue in Spanish: “Ayer fui al médico. ¿Qué le pasa? Me preguntó. Yo le dije: ¡Ay doctor! Donde pongo el dedo me duele. ¿Tiene la uña rota? Sí. Pues ya sabemos qué es. Deje su cheque a la salida.”
Classification Results of Spanish PD Subjects From the PC-GITA Dataset, Utilizing the Proposed Feature Sets and Combining all Using the Best-Performing Classifiers
As can be seen from Table 6 and Figure 11, the outcomes of our research hold significant clinical relevance, indicating an approximately 83% accuracy in classifying healthy controls versus individuals with PD Spanish speakers. Importantly, the balance between precision and recall suggests that our methodology does not exhibit bias towards either classification group. These high results reflect excellent discrimination capabilities, demonstrating that our comprehensive feature set—comprising 589 features that integrate acoustic features, Whisper embeddings, and transcription quality—effectively captures the speech patterns associated with PD. Depicting the classification results for spanish PD subjects from the PC-GITA dataset, utilizing the proposed feature sets and employing the best-performing classifiers
In our previous experiments, we utilized features extracted by Little et al from English-speaking individuals with PD 16 to classify Arabic PD speakers. Building on this foundation, we applied our own methodology, originally designed for Arabic PD speakers, to analyze Spanish PD speakers. The results were promising, demonstrating effective classification metrics across all three languages: English, Spanish, and Arabic. This consistency suggests that there are universal characteristics in the speech patterns of individuals with PD, regardless of the language spoken.33,34,82-85
These findings have major significance since they show that our approach not only works well across diverse linguistic settings, but also validates the idea that specific speech patterns linked with Parkinson’s disease may transcend language borders. Our work demonstrates the possibility for building globally applicable instruments for early diagnosis and assessment of Parkinson’s disease in various populations by attaining robust performance metrics across multiple languages. This study emphasizes the need of a multi-modal technique that can adapt to different languages while retaining high accuracy, which is critical for expanding clinical screening methods worldwide.
The 90% sensitivity and specificity achieved with LDA demonstrate the clinical utility of our approach, implying that it has potential for real-world screening. The cross-linguistic validation’s 83% accuracy on Spanish speakers indicates broad clinical applicability beyond Arabic populations.
SHAP Analysis
To understand and explain the effect of each feature on predicting PD based on our data, we used both the TF set and A2T excluding the deep features because they are not explainable by nature. For this purpose, we employed the Shapley Additive exPlanations (SHAP), a method proposed by Lundberg and Lee
86
to evaluate the importance of each feature and its effect on the trained model’s result. This technique applies game theory concepts and Shapley values to a variety of machine learning model interpretation methods. According to our previous section, the best performers were LDA and LinearSVC, both performed better on either dataset, therefore we merged both TF set and A2T in one dataset for PD prediction. Figure 12 illustrates the SHAP value results obtained by both models on the merged data. SHAP value results obtained by LDA and LinearSVC showing the most important feature in the decision-making process
The MFCCs are essential to both classifiers, as seen in Figure 11, with MFCC4, MFCC7, and MFCC21 continuously exhibiting a notable impact. Positive SHAP values are typically produced by higher MFCC values (shown in red). Additionally, the analysis highlights the importance of the lowest and median no-speech probability.
Figure 12 also shows important features such as the minimum and median compression ratios. Positive SHAP values are linked to higher compression ratios reflect the irregular, less predictable speech patterns characteristic of hypokinetic dysarthria, where normal prosodic flow is disrupted by motor control deficits. Pitch Mean shows a significant impact on both models; higher pitch variability (shown in red) leads to higher SHAP values. Additionally, the lower articulation rate and longer pauses, as measured by total pause duration and articulation rate, are further features of Parkinson’s speech. The SHAP value for the diagnosis of PD is improved by longer pauses and slower articulation rates (shown in blue).
As a result, TF set features like MFCCs, pitch, and articulation rate are good at capturing acoustic fluctuations, but A2T features (like log probabilities, no-speech probabilities, and compression ratios) offer more contextual information about fluency and coherence. The inability of Parkinson’s patients to speak continuously and fluently is best illustrated by whisper-derived metrics, specifically the sum of no-speech probabilities and average log probabilities.
The SHAP analysis finds patterns consistent with established PD speech pathology. The prevalence of pitch variability characteristics correlates with vocal fold stiffness produced by dopamine depletion in the basal ganglia. Similarly, the relevance of pause length and articulation rate metrics suggests bradykinesia, which impairs speech motor control in Parkinson’s disease patients. These SHAP-identified indicators might aid clinical evaluation by emphasizing the precise speech aspects most impaired in individual patients, thereby directing tailored speech therapy approaches.
Discussion
The use of ML technologies to identify particular voice pathologies is growing, particularly when it comes to providing a non-invasive approach and easily accessible early detection for PD, since some people may not even seek medical attention when the disease first appears and will likely go undiagnosed. According to research on speech and auditory perception, PD patients’ voices differ greatly. 87 However, a safe, easy, and economical approach is necessary for a reliable diagnosis of voice disorder. Hypokinetic dysarthria should be taken into account when using ML to identify PD. The results of this investigation, expand on earlier ML studies. This was achieved by showing how an ML method could be applied to detect instances of the dysarthria profile linked to PD, allowing for an accurate and prompt diagnosis.
The clinical implications of our findings have immediate translational potential. The high accuracy of 90% and the interpretable features identified through SHAP analysis lay a solid foundation for clinical decision support systems. Additionally, the cross-linguistic validation indicates that core speech biomarkers for Parkinson’s disease may be universal, allowing for the adaptation of our methodology to other linguistic populations with minimal modifications.
We evaluated the performance of linear ML classifiers in addition to various classification schemas, such as neural networks and decision trees, to more accurately diagnose PD using voice data. Based on the application scores, the top three linear classifiers, LDA, and Linear SVC, showed noticeably superior predictive power compared to the other methods. This can be explained by the linear link between PD progressions and speech feature anomalies, the most basic of which is the interplay between acoustic fluctuations, fluency, and coherence.29,88,89 The application of the SHAP analysis has further supported this result. On the other hand, earlier studies claimed that PD prediction is more precise when applied neural network methodologies. This result was following,21,22,90 and when using an additional deep learning approach in 20. It has been suggested that neural networks are perfect for big datasets because of their capacity to recognize intricate speech patterns. Neural networks, however, have higher computational requirements and complexity. Furthermore, the linear association between increased speech disability and the probability of PD along the correlation curve may not be taken into account by the non-linear approach at the extremes of PD cases. Yet, one important factor to take into account while using the method in clinical practice is its simplicity. 91 Moreover, in research involving patient cohorts, it is essential to maximize resources for collecting clinical data. We have a modest sample size in this study. Consequently, it makes sense to employ a predictive equation that can learn from small datasets. Since linear classifiers are easy to use and effective for linear connections and smaller datasets - a scenario that frequently arises in clinical studies - we are encouraging the adoption of the presented ML pipeline by other groups tackling similar problems.
We acknowledge that the high feature-to-sample ratio (589 features vs. 40 subjects) poses an inherent risk of overfitting, which warrants careful consideration. Several design choices in our methodology mitigate this risk, these include. 1- The LOO-CV maximizes training data in each fold (39 out of 40 samples), providing a nearly unbiased performance estimate for our small dataset. 2- The PCA reduced the 589 features to 10 principal components preserving 99.49% of variance, substantially lowering the effective dimensionality. 3- The RF-based feature selection reduced the feature set to 57, resulting in a more favorable feature-to-sample ratio. 4- The choice of LDA and LinearSVC as top-performing classifiers is itself a form of regularization, as these linear models have limited capacity to memorize noise compared to nonlinear alternatives. 5- Most importantly, the successful external validation on the PC-GITA Spanish dataset (100 subjects, ∼83% accuracy) provides a strong evidence that our models capture genuine PD-related speech patterns rather than dataset-specific artifacts.
Nevertheless, we recognize that the small sample size remains the major limitation of our study, and therefore, future work should aim to replicate these findings on larger Arabic-speaking cohorts.
Our study emphasized many methodological aspects of speech analysis. Instead of using a preexisting dataset, as in other investigations,20,21 we established a new PD cohort focused on native Arabic speakers. The data collection process of PD patients presented several challenges, similar to any cohort study. These challenges include poor patient motivation and inconsistent or missing address and contact information. Many patients were unable to participate due to their advanced conditions, such as significant disabilities or the inability to walk or stand without assistance. This explains why, compared to previous studies, our dataset seems to be relatively small. Despite this, we were able to obtain valuable information on how the feature extraction strategy and the number of accumulated features affect the effectiveness of predictive diagnosis based on ML.
Our findings highlight the significance of each feature in training the ML classifiers, particularly when using the LDA model. However, a good performance of the Linear SVC model was achieved when leveraging the feature selection results, utilizing up to 57 features rather than 589. Alshammari et al. 20 have emphasized the significance of using every feature in the training process, which is consistent with our findings. Nevertheless, applying feature selection in their dataset did not produce acceptable outcomes. It is anticipated that different ML classifiers will exhibit varying degrees of performance in various data situations. Possible explanation for this outcome is that the feature-ranking algorithm estimates the importance of a feature’s correlation with the class while depreciating the inter-correlations of features. Based on these suppositions, the primary distinctions between the two approaches may be summarized as follows: the statistical markers utilized for calculating each correlation, and the various search techniques used (a ranker suggesting a single-feature selection process). The risk of redundant sets arises as information gain, in contrast, does not take into account the inter-correlations between features and, instead, is predicated on the amount of information learned about a random variable (class) from observing another random variable (feature). 18 However, as feature selection minimizes the model size and computation time, we recommend utilizing the feature selection approach, particularly when the dataset is much bigger than ours be.
The comparative accuracy results of the feature selection vs. all feature strategies, which does not necessitate such meticulous attention to detail or the selection of features and extraction methods, make the benefits of a linear-based approach evident. These benefits include less formal statistical training, the capacity to implicitly detect relationships between dependent and independent variables, and the ability to identify all potential interactions between predictor variables. Nevertheless, modification and augmentation of data before testing, and optimization algorithms, are still important aspects in this context.
The most relevant voice features selected by RF included those that have already been discovered in studies on voice analysis for PD: basic frequency features such as F0, shimmer, and jitter can be summarized as the most frequently used acoustic features, in addition to the MFCC features. This is evident from looking at the illustrated summary of the highest-ranked attributes provided in Figure 5. This outcome is consistent with the pertinent literature and supports the biological validity of our findings.14-17,19,21
Moreover, the SHAP analysis in this work enables zooming in on specific predictions and offers a global sense of feature importance. The influence of the predictor variables and their interactions in the suggested pipeline are examined in this study by presenting the LDA and Linear SVC using SHAP analysis. The model’s predictor variables’ behavior was better understood according to the SHAP. MFCC coefficients (MFCC4, MFCC7, MFCC21) capture formant frequency changes consistent with the vocal tract rigidity observed in PD patients, where reduced articulatory precision affects resonant frequencies. This finding aligns with recognized speech difficulties associated with PD, including alterations in vocal tract resonance. 92 The analysis also emphasizes how crucial the lowest and median no-speech probabilities are. Prolonged speech gaps are less common in healthy people, PD patients, on the other hand, frequently have fragmented speech patterns, which are reflected in greater no-speech probabilities. 93
Moreover, the SHAP analysis highlights the importance of the compression ratios indicating that less compressible speech, which reflects jerky or irregular patterns, predicts PD. Higher pitch variability leads to higher SHAP values. The predictive character of pitch changes can be explained by the more monotonous or tremulous speech that Parkinson’s disease patients frequently exhibit. 87 Additionally, the lower articulation rate and longer pauses, as measured by total pause duration and articulation rate, are further features of Parkinson’s speech. 93
Clinical research has shown that people with PD often deviate more from healthy people in voice recordings, especially when it comes to more lengthy pronunciation measures. 94 Furthermore, linguistic differences between individuals of different races may potentially lead to subpar classification results. 95 Nevertheless, using a pre-trained Whisper model significantly improved the dysarthria-based PD classification outcomes in this study. The results of this study also highlighted the usefulness of specific vocal qualities, which is consistent with previous research on voice analysis for Parkinson’s disease. Notably, using a pre-trained Whisper model enhanced the PD classification results. 58
Using the dimensionality reduction linear transformation, the PCA improved the classification outcome in almost all tested algorithms, compared to the non-linear transformations (SEM and t-SNE), which fared poorly across all metrics except in the case of the AdaBoost algorithm, which performed well on all features. However, none of these approaches compared favorably to the feature selection or “all features” approaches. Prior research has highlighted the advantages of these algorithms, such as their capacity to improve task outcome classification. 27 PCA searches for features that show the most variance between classes to construct the principal component space. Utilizing the concepts of variance matrix, covariance matrix, eigenvector, and eigenvalue pairs, the technique performs PCA and produces a set of eigenvectors and their corresponding eigenvalues.
The shared underlying dynamics between t-SNE and SEM explain the similarities between their outcomes. It appears that distinct traits are identified as being most significant by the less effective t-SNE. The results indicated that the majority of features found using SEM and t-SNE were associated with a feature type that is indicative of a low classification level or which is mostly constituted of high-level differential features (e.g., MFCC). The trends in the top ranked features tentatively validate how pitch-related and prosodic features contain pertinent information for the detection and staging of PD using voice data, which is supported by the observation that the RF seemed to be the best-performing feature selector.
This work opens the door for more research to compare the dysarthria-based approach with other biomarkers, such as Laryngeal Imaging Biomarkers, 96 Aerodynamic Measures, Electrophysiological Measures, 97 Genetic Biomarkers, Inflammatory Biomarkers, Psychosocial Factors, Neuromuscular Biomarkers, 98 Metabolomic Profiles, and Vocal Fold Histopathology. 99 Incorporating a combination of these biomarkers can improve diagnostic accuracy and facilitate understanding of the multifactorial nature of dysarthria.
Conclusion
We proposed a system that is resilient to fluctuations, enabling it to derive significant features that remain constant despite the extensive range of speech variations in dysarthria. This was made evident through a speech dataset acquired for our research, effectively capturing the distinction between PD and HC speech at different intensities. Furthermore, through employing machine learning, the model may be trained to extract high-level features that capture crucial speech properties that are significant for dysarthria-based PD classification.
Furthermore, the proposed ML approach, which employs a cutting-edge pre-trained speech recognition model (Whisper), MFCC, and fundamental frequency features, as well as RF Feature Selection and PCA dimensionality reduction, performs well, especially when using the LDA and LinearSVC classifiers, where the accuracy measures of PD identification approaches to 90%, this result is achieved using all the extracted features from the three proposed sets of features.
Moreover, the proposed technique does not require sophisticated computer hardware and may be easily applied as an online tool for medical applications. The proposed ML methodology was validated against state-of-the-art model-based features (Whisper), MFCC, and basic frequency features, as well as by comparing many ML pipelines and analyzing their precision. We were unable to assess the performance of the established models on a larger dataset—for example, considering more PD stages or medication status—due to resource constraints.
This work establishes the first validated Arabic PD speech classification system and demonstrates cross-linguistic generalizability, providing a foundation for the global implementation of speech-based PD screening. The interpretability of the methodology through SHAP analysis facilitates clinical translation, while the universal features identified suggest potential for broader applications across different languages.
Our future work will be guided to provide solid solutions for the major limitations of this study, which include the small size of the dataset. Also, more comparison studies will be conducted with other sophisticated techniques on the same dataset in future studies. Although our sample size is modest, the rigorous cross-validation approach and successful external validation help alleviate concerns about overfitting. A limitation of our study is the lack of detailed medication timing data, which future research should aim to address; however, it is worth noting that all patients were on stable regimens during the recording.
Footnotes
Acknowledgments
We would like to extend our heartfelt gratitude to Juan Rafael Orozco Arroyave, Full Professor at the GITA Lab, Universidad de Antioquia (UdeA), and his research center for generously providing access to the dataset used in this study. Their contribution has been invaluable in advancing our research on Parkinson’s disease classification, and we appreciate their commitment to fostering collaboration within the scientific community. Access to this dataset has significantly enriched our analysis and enabled us to draw meaningful conclusions from our findings. Additionally, the first author sincerely thanks the Deanship of Scientific Research at Mutah University for its financial assistance, which is identified by grant number 757/2023.
Ethical Considerations
The study received ethical approval from the institutional ethics committee at the University of Mutah (17/89/871), in addition to another approval obtained by the Ministry of Health, Jordan (number 11936). All participants provided written informed consent prior to their inclusion in the study, and the participant demographic and clinical features, including the audio recordings of the patient and control group, were documented anonymously. The patients were accessed in collaboration with the neurological departments of Al-Karak Governmental Hospital, Karak, Jordan, and King Abdallah Hospital, Irbid, Jordan.
Consent for Publication
No individual patient data or identifiable images are presented in this manuscript. All audio recordings and demographic data were collected and stored anonymously.
Author Contributions
All authors contributed equally to this work and approved the final manuscript.
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Data Availability Statement
The voice dataset that supports the findings of this study was recorded by different subjects in different areas in Jordan, and it is available from the corresponding author upon reasonable request.