A Comparative and Explainable Study of Machine Learning Models

for Early Detection of Parkinson's Disease Using Spectrograms

Hadjer Zebidi

, Zeineb BenMessaoud

and Mondher Frikha

Advanced Technologies for Image and Signal Processing ‘ATISP’ Research Lab,

National School of Electronics and Telecommunications of Sfax (ENET’Com), University of Sfax, Sfax, Tunisia

Higher Institute of Computer Science and Multimedia of Gabes (ISIMG), Gabes University, Gabes, Tunisia

Keywords: Parkinson’s Disease (PD), Early Detection, Spectrogram, Machine Learning, SMOTE, GridSearchCV, LIME.

Abstract: Parkinson's disease (PD) is a progressive neurodegenerative disorder that originally affects the motor system.

Therefore, early diagnosis is essential for effective intervention. Classic diagnostic approaches heavily rely

on clinical observations and manual feature extraction, limiting the detection of subtle early vocal

impairments. This research examines machine learning (ML) techniques, namely Support Vector Machines

(SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost), for early identification of PD

through the analysis of spectrogram images derived from voice recordings. Mel-Frequency Cepstral

Coefficients (MFCC), Short-Time Fourier Transform (STFT), and Mel-Spectrograms were extracted. The

improvement of the model was introduced by the Synthetic Minority Over-sampling Technique (SMOTE)

and hyperparameter tuning using GridSearchCV (Grid Search with Cross-Validation). Implementing the

above methods resulted in significant performance improvements, with XGBoost achieving an accuracy of

95 ± 0.02 on the PC-GITA dataset and SVM attaining 90.74 ± 0.04 on the Neurovoz dataset. Local

Interpretable Model-agnostic Explanations (LIME) enhanced model transparency by identifying the

significant regions in spectrograms that most influence predictions. This analysis illustrates the efficacy of

ML models utilizing SMOTE and GridSearchCV, particularly when augmented by LIME for interpretability,

in improving early detection of PD, thereby presenting a feasible approach for clinical implementation.

1 INTRODUCTION

Parkinson's disease, or PD, is a neurodegenerative

illness that affects about 1% of people over 60 around

the world (Dorsey et al., 2018). The condition is

marked by a group of motor symptoms, such as

tremors, rigidity, and slow movement, as well as a

number of non-motor symptoms, such as cognitive

decline and changes in speech and voice (Bloem et

al., 2021). These symptoms significantly affect the

quality of life for patients with PD and provide

considerable obstacles for healthcare personnel in the

prompt and efficient administration of medication.

Early PD diagnosis is paramount, as it allows for

implementing therapeutic interventions that can

markedly enhance patient outcome (Murman, 2012).

Still, the present diagnostic techniques—which

mostly rely on clinical assessments and patient-

https://orcid.org/0009-0003-8204-9712

https://orcid.org/0000-0003-2584-5141

reported symptoms—often insufficient short for early

PD (Gullapalli & Mittal, 2022), particularly in cases

of subtle or unrecognized voice impairments.

People with PD show particular changes in their

vocal features, including a decrease in pitch

variability, changes in speech pace, and articulation

problems (Harel et al., 2004). While audio recordings

allow one to record these changes, conventional

analysis techniques may rely on hand feature

extraction, which might not fully capture the intricacy

of voice patterns, therefore restricting their efficacy

in automated assessments (Klempíř & Krupička,

2024).

Due to technological developments, speech

analysis— as a non-invasive and easily available

technique for early disease diagnosis—has been

developed. Analyzing acoustic features including

pitch, jitter, and shimmer shows that it is possible to

272

Zebidi, H., BenMessaoud, Z. and Frikha, M.

A Comparative and Explainable Study of Machine Learning Models for Early Detection of Parkinson’s Disease Using Spectrograms.

DOI: 10.5220/0013183900003905

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 272-282

ISBN: 978-989-758-730-6; ISSN: 2184-4313

distinguish PD patients from healthy people.

However, many research using these features depend

on traditional ML models, which need for hand

feature engineering and might not be sufficient in

capturing the whole spectrum of auditory features

(Badhan & Kaur, 2024).

Building on these developments, researchers aim

to leverage the strengths of ML methods to assess

their performance in identifying early PD signs using

spectrogram images of voice recordings. Optimal for

ML-based analysis, spectrograms graphically depict

audio signals and capture both frequency and time-

domain information.

Using the PC-GITA and Neurovoz datasets, this

work contrasts the performance of several ML

models—including SVM, RF, and XGBoost—to

identify early-stage PD from continuous vowel

recordings of the vowel "a". We used SMOTE and

GridSearchCV hyperparameter adjustment to solve

class imbalance and optimize model performance.

To enhance model interpretability, we employed

LIME in the best-performing ML models. This method

helps to distinguish how ML models interpret speech

data for early PD detection and helps to identify which

areas of the spectrograms most contributed to the

model

's capacity. This understanding is crucial for

identifying the vocal features that separate PD patients

from healthy controls (HC), thus validating the model's

predictive capabilities.

This work attempts to evaluate, using voice

analysis, whether ML techniques are more effective

for early PD identification. This comparison

underscores the potential of ML techniques could

automatically extract complex properties.

The present paper focuses on:

• The utility of spectrogram images for

classification purposes.

• A comparative analysis of ML techniques

for the early detection of PD.

• Model interpretability through LIME.

The framework of this research is presented in the

next parts. With a focus on spectrogram production,

feature extraction, and model training, Section 2

presents the framework for the materials and

methodology. Section 3 offers a comprehensive

review of the evaluation metrics critical for dataset

assessment. Last but not least, Section 4 presents a

thorough examination of the results, contrasting the

performance of ML models and stressing important

results; it also includes the conclusions of the study

and suggestions for next studies.

2 MATERIALS AND METHODS

This paper illustrates the proposed framework, which

employs a systematic methodology for early

classifying PD based on speech datasets, as shown in

Figure 1. It entails gathering voice data from patients

with PD and HC participants for analysis, followed

by data preprocessing. The feature extraction step

commences, extracting attributes such as MFCC,

STFT spectrogram, and Mel-spectrogram from the

audio data, converting them into spectrogram images,

and balancing the dataset utilizing SMOTE.

Following data balancing, the next step is to employ

GridSearchCV to refine the hyperparameters of the

ML models and evaluate their efficacy. We organize

each phase to guarantee that the models can

proficiently discriminate between HC and individuals

with PD utilizing speech data, therefore facilitating

the early detection and diagnosis of the disease.

2.1 Parkinson’s Disease Dataset

Two extensively referenced (PD) datasets were

employed in this study: PC-GITA (Orozco-Arroyave

& Noth, n.d.) and Neurovoz (Mendes-Laureano et

al., 2024). Each database comprises individuals with

PD and HC subjects. Neurologists have employed the

Unified Parkinson's Disease Rating Scale (UPDRS)

and the Hoehn and Yahr scale (H&Y) to identify and

categorize the Patients. The databases exhibit

variations in demographics and sizes, as outlined in

Table 1.

All recordings utilized in this study were at an

early stage (UPDRS stages 1-2). Under controlled

environmental settings, data recordings were

conducted in all instances. The staff instructed each

participant to execute several speech activities. This

study examines sustained phonation of the vowel /a/.

Each patient was recorded three times.

Table 1: Demographic information, including gender and age ranges for the PC-GITA and Neurovoz corpora.

Corpus

Subjects Age (Years)

Female Male Female Male

PD HC PD HC PD HC PD HC

Neurovoz 16 23 15 24 56–86 58–86 41–80 53–77

PC-GITA 17 25 19 25 44–75 43–76 33–77 31–86

A Comparative and Explainable Study of Machine Learning Models for Early Detection of Parkinson’s Disease Using Spectrograms

273

Figure 1: Proposed model for early PD classification.

2.2 Data Pre-Processing

All recordings from the PC-GITA and Neurovoz

datasets were resampled to 16 kHz utilizing the

librosa library to maintain uniformity in the sampling

rate. EBU R128 loudness normalization was

implemented utilizing the ffmpeg-normalize library

(Ffmpeg-Normalize, n.d.) to attain uniform loudness

levels, enhancing conventional peak-based

normalization.

The spectrogram features were retrieved and

saved as 224x224-pixel image files. These images

were organized by class label (HC for healthy

controls, PD for Parkinson's disease) then normalized

to the range [0, 1] by dividing pixel values by 255.

80% of the dataset was used for training, and 20%

was used for testing. The dataset was divided into

training and test sets at random.

2.3 Feature Extraction

2.3.1 MFCC Coefficients

MFCCs were extracted with 13 MFCC coefficients

per frame and their derivatives in order to capture the

envelope of a sound's short-term power spectrum

(Mishra et al., 2024). These factors are perfect for

differentiating between PD patients and HC because

they accurately capture the timbral features of speech.

2.3.2 STFT Spectrograms

The STFT spectrogram is an effective tool for

analyzing and visualizing time-varying frequency

content in audio signals. It converts all recordings into

the time-frequency domain, facilitating the

examination of the dynamic evolution of frequency

content in a given signal (Xuan, 2023). In this work,

windows (n-fft) of 32 milliseconds in length were

computed using STFT representations chosen to

achieve a balance between computing efficiency and

sufficient temporal and frequency precision. To

ensure constant signal length, the hop length was set

at 8 milliseconds and the maximum padding length

(max_pad_len) was changed to 100. The time-

frequency representation extracted was used as input

to feed to the ML model.

a. HC speaker b. PD patient

Figure 2: STFT spectrograms of the vowel /a/speech signal

pronounced by HC speaker (a) and a patient with PD (b).

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

274

2.3.3 Mel-Spectrograms

The Mel-spectrogram is a popular used representation

of audio data that measures the power of a signal on

the Mel scale, which more accurately approximates

the human auditory system's reaction to different

frequencies. This work created Mel-spectrograms

with a filter bank of 40 Mel bands. This method

transforms the power spectrogram to a log scale,

resulting in a reliable representation of audio features

relevant for speech analysis and disease detection.

After extracting features from the MFCC with

their derivatives, STFT and Mel-spectrograms, the

spectrograms were flattened into one-dimensional

vectors, thus ensuring compatibility with ML models.

In this stage, the two-dimensional spectrograms are

converted into a format that numerical array-input

models may use. A full feature set representing each

audio sample was created by merging the resulting

flattened vectors. To make sure that all relevant

auditory features were used for classification, these

feature vectors were fed into the ML pipeline.

2.4 Classification Methods

Following previous stages, we used ML techniques,

including SVM, RF, and XGBoost, to identify early

PD, building on earlier stages. To improve the

performance of the model and find the ideal

hyperparameters, SMOTE and GridsearchCV were

used to all of the ML models. Kernel types, the

number of estimators, and learning rates were

explored to determine the most effective models for

the early detection of PD.

2.4.1 Support Vector Machine (SVM)

A supervised ML algorithm that is widely utilized for

classification tasks. It works by finding a hyperplane

that optimally differentiates data points belonging to

disparate classes within a high-dimensional space.

SVM is efficacious in high-dimensional settings and

exhibits resilience to overfitting, particularly when

the number of features surpasses the number of

observations. Employing an array of kernel functions

(linear, polynomial, radial basis function) can handle

both linear and non-linear classification (Bind et al.,

2015).

2.4.2 Random Forest (RF)

An ensemble learning method based on decision

trees. During the training phase, it builds multiple

trees, each using a random subset of the data. The

outputs of these trees are then combined using

majority voting, resulting in an improved

classification. The method is versatile, capable of

handling large datasets with high dimensionality, and

resistant to noise and overfitting (Breiman, 2001).

2.4.3 Extreme Gradient Boosting (XGBoost)

A robust gradient-boosting algorithm that builds

models stage-wise, with each new tree correcting

errors from the previous ones by focusing on

misclassified examples. It uses regularization to

minimize overfitting and use optimization techniques

such as parallelization and tree pruning to improve

performance and speed. XGBoost excels at processing

structured data and consistently outperforms other

algorithms in ML contests (Wang et al., 2022).

2.4.4 Synthetic Minority Over-sampling

Technique (SMOTE)

There are few PD samples than HC samples in the

dataset, which creates an imbalance. A solution to this

problem is to oversample members of the minority

class in order to achieve distributional parity.

Improving the class distribution is possible via

instance replication, but it does not provide any new

information. SMOTE addresses this issue by

producing new samples using linear interpolation of

existing minority class instances, hence generating

synthetic data points along the trajectory of the

feature space (Brownlee, 2020). For a certain

minority instance 𝑥



a synthetic sample 𝑥

𝒏𝒆𝒘

generated by interpolating between 𝑥

𝒊

and one of its

k-nearest neighbors 𝑥

𝐧𝐞𝐢𝐠𝐡𝐛𝐨𝐫

, where λ is a random

value within the interval [0,1].

𝑥

𝒏𝒆𝒘

=𝑥



+𝝀(𝑥

𝐧𝐞𝐢

𝐠

𝐡𝐛𝐨𝐫

−𝑥

𝒊

) (1)

2.4.5 Hyperparameter Tuning

(GridSearchCV)

GridSearchCV is a widely utilized method for

hyperparameter optimization in ML models.

Hyperparameters are predefined configurations set

prior to model training, encompassing the Kernel, C,

gamma, number of estimators, learning rate, and

max_depth. Determining the ideal values for these

parameters can substantially enhance model

performance (Jumanto et al., 2024).

GridSearchCV is a method for systematically

exploring a predefined set of hyperparameter values

to identify the optimal configuration. It requires three

key elements:

 The estimator represents the model to be

trained.

A Comparative and Explainable Study of Machine Learning Models for Early Detection of Parkinson’s Disease Using Spectrograms

275

 A parameter grid is a list of hyperparameters

and their potential values.

 CV: The number of folds in K-fold cross-

validation. We set k = 5.

3 EVALUATION METRICS

The key evaluation metrics for classification models

are accuracy, precision, recall, and F1-score:

 Accuracy measures the proportion of correctly

classified instances:

Accurac

𝑇𝑃 + 𝑇𝑁

𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁

(2)

 Precision indicates the proportion of predicted

positives that are positive:

Precision =

𝑇𝑃

𝑇𝑃 + 𝐹𝑃

(3)

 Recall measures the proportion of actual

positives correctly identified:

Recall =

𝑇𝑃

𝑇𝑃 + 𝐹𝑁

(4)

 F1-score balances precision and recall,

particularly for imbalanced datasets:

F1 − score =

2 𝑇𝑃

2 𝑇𝑃 + 𝐹𝑃 + 𝐹𝑁

(5)

4 RESULTS AND DISCUSSION

Tables 2 and 3 present a comparative performance of

ML models for early PD detection from the two

disjoint datasets without and with the application of

SMOTE. The analysis is based on performance

metrics, including accuracy, F1 score, precision, and

recall, all with standard deviation, to highlight the

strengths and weaknesses of the models across these

datasets.

The models showed different success over the

two datasets according to the results acquired

without applying SMOTE (Table 2). The XGBoost

model showed the best level of accuracy in the PC-

GITA dataset—82.85 ± 0.05 Followed by the SVM

(80.77 ± 0.05). While the XGBoost model showed

the highest precision (86.67 ± 0.08), the SVM model

also had the highest F1-score (77.27 ± 0.08) and

recall (73.42 ± 0.09). Its F1-score and recall were,

however, rather lower than those of SVM and RF.

The results imply that XGBoost shows a superior

balance between accuracy and precision. With an

accuracy of 78.85 ± 0.07 and a recall of 59.09 ± 0.14

RF model showed the lowest performance among

the three. The results show that the RF model

showed more trouble than the other models handling

class imbalance.

With an accuracy of 72.09 ± 0.06, an F1 score of

70 ± 0.04, and a recall of 66.67 ± 0.12, SVM displays

the best performance in the Neurovoz dataset. With

XGBoost attaining the lowest accuracy (69.37 ± 0.04)

and recall (52.38 ± 0.10), the results for XGBoost and

RF were inferior; yet, its precision was greater than

those of other models with 78.57 ± 0.04. These results

show that although SVM showed superior

performance in handling data without SMOTE,

XGBoost, and RF showed more amazing difficulty in

addressing the class imbalance, especially with the

recall.

Table 2: Performance results without SMOTE from each dataset using the ML models for the early prediction of PD.

Dataset Model Accuracy F1_score Precision Recall

PC-GITA

SVM 80.77 ± 0.05 77.27 ± 0.08 79.86 ± 0.11 73.42 ± 0.09

RF 78.85 ± 0.07 70.27 ± 0.13 85.91 ± 0.09 59.09 ± 0.14

XGBoost 82.85 ± 0.04 70.27 ± 0.08 86.67 ± 0.08 59.68 ± 0.06

Neurovoz

SVM 72.09 ± 0.06 70 ± 0.04 73.68 ± 0.03 66.67 ± 0.08

RF 70.79 ± 0.05 66.67 ±0.05 72.22 ± 0.07 61.90 ±0.12

XGBoost 69.37 ± 0.04 62.86 ± 0.07 78.57 ± 0.04 52.38 ± 0.10

Table 3: Performance results with SMOTE from each dataset using the ML models for the early prediction of PD.

Dataset Model Accuracy F1_score Precision Recall

PC-GITA

SVM 90 ± 0.03 89.60 ± 0.04 96.30 ± 0.05 83.87 ± 0.06

RF 86.66 ± 0.05 86.14 ± 0.06 96 ± 0.03 77.42 ± 0.07

XGBoost 95 ± 0.02 95.08 ± 0.03 96.67 ± 0.02 93.55 ± 0.04

Neurovoz

SVM 90.74 ± 0.04 90.71 ± 0.05 96.15 ± 0.04 86.21 ± 0.05

RF 83.33 ± 0.06 84.21 ± 0.07 85.71 ± 0.05 82.76 ± 0.06

XGBoost 80.62 ± 0.07 81.36 ± 0.06 80 ± 0.08 82.76 ± 0.03

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

276

When SMOTE (Table 3) was used, the

performance of all models improved, especially SVM

and XGBoost, across both datasets. When tested on

the PC-GITA dataset, the SVM was enhanced. Its

accuracy went from 80.77 ± 0.05 to 90 ± 0.03 and its

F1 score went from 77.27 ± 0.08 to 89.60 ± 0.04. The

accuracy went up a lot, from 79.86 ± 0.08 to 96.30 ±

0.05, and the recall went up a lot, too, from 73.42 ±

0.09 to 83.87 ± 0.06. It's clear from these results that

SMOTE facilitated enhanced the SVM's performance

by reducing class mismatch, which is a key part of

finding rare PD cases in the dataset.

XGBoost significantly improved performance,

with accuracy increasing from 82.85 ± 0.04 to 95 ±

0.02, F1 score went from 70.27 ± 0.08 to 95.08 ± 0.03,

and recall went from 59.68 ± 0.06 to 93.55 ± 0.04.

Based on these results, it looks like the synthetic data

that SMOTE created helped XGBoost a lot,

especially when it came to memory. This means that

the model became better at finding positive cases of

PD. Even though RF's accuracy went up from 78.85

± 0.07 to 86.66 ± 0.05, its F1 score and recall went up

less than those of SVM and XGBoost. As expected,

the recall went up from 59.09 ± 0.14 to 77.42 ± 0.07.

This indicates that SMOTE enhanced performance

for RF, albeit not to the same extent as for SVM and

XGBoost.

The enhancements on the Neurovoz dataset were

more modest but still significant. The SVM's recall

improved from 66.67 ± 0.08 to 86.21 ± 0.05, and its

accuracy increased from 72.09 ± 0.06 to 90.74 ± 0.04.

These results indicate that SVM experienced

substantial benefits from SMOTE, as evidenced by its

improved generalization across the entire dataset and

significant increases in recall. In addition, XGBoost

exhibited an increase, with accuracy increasing from

69.37 ± 0.04 to 80.62 ± 0.07 and recall increasing

from 52.38 ± 0.10 to 82.76 ± 0.03. However, the

model's susceptibility to the challenge was suggested

by the fact that the XGBoost enhancement was less

pronounced on Neurovoz than on PC-GITA.

Figure 3 illustrates the ROC-AUC curves of the

three ML models—SVM, RF, and XGBoost—

generated on the PC-GITA dataset. In subplot (a), the

AUCs for SVM and XGBoost are 0.90 and 0.88,

respectively, indicating excellent discriminatory

performance without SMOTE. Nevertheless, the RF

model exhibits slightly superior predictive capacity,

with an AUC of 0.90. These findings show that all

three models can efficiently identify PD patients from

HC, with RF doing the best without SMOTE. In

subplot (b), the use of SMOTE improved the

performance of all three models. The SVM and

XGBoost models both have a good AUC of 0.96,

suggesting exceptional classification skill across

thresholds.

Moreover, RF has superior performance, with an

AUC of 0.91, highlighting its capacity to tackle class

imbalance via SMOTE. The observed enhancement

across models demonstrates that oversampling using

SMOTE significantly improves the model's

robustness, especially in datasets with class

imbalance. Both SVM and XGBoost exhibit

exceptional and consistent performance on the PC-

GITA dataset, rendering them optimal choices for

early PD identification.

Figure 3: AUC-ROC curve of the three ML models in PC-GITA dataset without and with SMOTE.

(

)

Without SMOTE

(

)

With SMOTE

A Comparative and Explainable Study of Machine Learning Models for Early Detection of Parkinson’s Disease Using Spectrograms

277

Figure 4: AUC-ROC curve of the three ML models in Neurovoz dataset without and with SMOTE.

Figure 4 demonstrates the effectiveness of the

models on the Neurovoz dataset, as evidenced by the

AUC-ROC curves. In subplot (a), before to the

application of SMOTE, the SVM attains a maximum

AUC of 0.79, which, although satisfactory, indicates

a reduction in performance relative to PC-GITA.

XGBoost achieves an AUC of 0.73, whereas RF has

the lowest performance at 0.70. These findings

indicate that the Neurovoz dataset poses additional

hurdles, most likely due to variability and underlying

characteristics in patient data. Subplot (b) shows that

applying SMOTE improves AUC scores significantly

across all models, illustrating the relevance of

resolving class imbalance. SVM once again earns the

highest AUC (0.96), demonstrating its resilience and

reliability. The RF model shows a significant

improvement, with an AUC of 0.92, demonstrating

that it can adjust to class-balanced data. XGBoost has

an AUC of 0.89, which indicates improved

performance in this context.

The importance of SMOTE in addressing class

imbalance is emphasized by these findings,

particularly in the context of the Neurovoz dataset.

GridSearchCV was also instrumental in the

optimization of hyperparameters for all models,

which exacerbated the performance enhancements

observed with SMOTE. Furthermore, fine-tuning

parameters such as kernel type and regularization

strength improved SVM's capacity to generalize,

contributing to the model's consistently high AUC.

Similarly, with XGBoost, GridSearchCV enhanced

the learning rate and maximum tree depth, allowing

the model to catch more complicated patterns in the

data. After applying SMOTE, RF gained much

improved performance from optimal tree depth and

the number of estimators. These hyperparameter

adjustments highlighted the joint effectiveness of

SMOTE and GridSearchCV by helping the models

match their performance with the features of every

dataset. Several significant observations arise from

results analysis. The application of SMOTE,

illustrated in subplot (b), significantly enhanced AUC

values for all models and datasets, thereby addressing

class imbalance concerns. Second, SVM's

dependability for early PD identification was

confirmed by its consistent achievement of the

greatest AUC across datasets (0.96). XGBoost

demonstrated exceptional efficacy, particularly on

PC-GITA, despite minor performance fluctuations.

RF, despite its initial subpar performance,

successfully adjusted to SMOTE and achieved

impressive results on the Neurovoz dataset.

Nonetheless, SMOTE facilitated the recovery of all

models, resulting in robust outcomes.

In summary, the SVM model proves to be the

most resilient and reliable across both datasets and

situations, positioning it as a formidable candidate for

clinical applications. XGBoost and RF exhibit

potential, especially when customized to the distinct

features of certain datasets. The synergistic effect of

SMOTE and GridSearchCV underscores their

significance in improving model efficacy for

unbalanced datasets, especially with early PD

detection.

4.1 Local Interpretable

Model-Agnostic Explanations

(LIME)

In this study, we employed LIME to interpret the best

model, XGBoost, predictions in classifying PD in

early-stage and HC patients. A post-hoc

interpretability method was developed to explain

(

)

Without SMOTE

(

)

With SMOTE

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

278

every prediction complex and black-box model. It

achieves this by constructing local, interpretable

models that approximate the behavior of the black-

box ML models within some small neighborhood

around the data point being explained (Molnar, 2020).

LIME was selected for this work because it is a

model-agnostic technique. Hence, it can be easily

applied to any ML algorithm (e.g. XGBoost, SVM).

It will also provide local explanations that will enable

the analysis of exactly which features of the input

data- namely, Mel-spectrograms- the model is

considered most influential for classifying each

patient as either PD or HC. This capability is critical

when applied clinically, for which there is an urgent

need to understand the underlying decision-making

process of ML models to build trust among health

professionals and ensure the validity of model

outputs.

LIME works by perturbing the input data, slightly

modifying the spectrogram, and observing how the

model's prediction changes. Then, it builds a more

straightforward interpretable model example; a linear

model approximates the decision boundary of the

complex model in the vicinity of the input data point.

Formally, this can be presented as:

(x)

𝑎𝑟

𝑔

𝑚𝑖𝑛 L

(







+Ω

(



(6)

Where :

• f is the original black-box model (XGBoost in

our study).

• g is the interpretable local surrogate model.

•





is a proximity measure that assigns higher

weights to data points close to the instance

• L (f, g, π

) ensures that the local surrogate

model g approximates the complex model f

behavior.

• Ω(g) enforces interpretability by ensuring the

surrogate model remains simple.

In our analysis, LIME has been applied to explain

the predictions made through XGBoost upon

spectrograms from PC-GITA and Neurovoz datasets.

The spectrograms represent time-frequency

representations of speech recordings, where XGBoost

was tasked with classifying the spectrograms as

belonging to either PD or HC patients. With LIME,

we can visualize those regions within the Mel-

spectrograms that contribute the most toward model

predictions and, in turn, provide an interpretable

explanation for each decision made by this class.

Figures 5 and 6 present LIME explanations for PD

and HC patients of the PC-GITA and Neurovoz

datasets, respectively.

Figure 5 illustrates that the XGBoost model

utilized four essential regions inside the Mel-

spectrogram to categorize PD patients in the PC-

GITA dataset. The locations marked in red signify

areas where the model identified auditory traits

indicative of PD, including diminished frequency

variability, lower vocal intensity, and delayed speech.

These features signify vocal abnormalities typically

linked to PD, such as monotone speech and

dysarthria. The localized features of these

emphasized regions indicate that the model

concentrates on distinct portions of the spectrogram

to discern PD-specific speech patterns.

Furthermore, the model found five separate areas

that assisted with sort the PC-GITA dataset into

groups of HC patients. The red areas show the sound

features that show a healthy vocal system. These

include changing pitches quickly and clearly, and

regular speech. These traits show normal speaking

range and flow, which is very different from the stiff

and repetitive speech patterns seen in people with PD.

Focusing on these key areas shows that the model

uses variations in frequency and intensity to tell the

difference between normal and Parkinsonian

speaking.

Figure 5: LIME explanations for both PD and HC from PC-

GITA dataset.

A Comparative and Explainable Study of Machine Learning Models for Early Detection of Parkinson’s Disease Using Spectrograms

279

By contrast, in the case of the Neurovoz dataset,

XGBoost depended on a broader set of regions in both

PD and HC Mel-spectrograms. In the case of the HC

patients from this dataset, three regions were

underlined by the model within the spectrogram. This

wider distribution of salient regions reflects the

greater complexity and variability inherent in healthy

speech patterns. Healthy subjects thus had dynamic

pitch modulation, more variance over time, and a

wider range of vocal frequencies, while the

dependence of the model on several regions indicates

that this is a level at which complexity needs to be

captured for accurate classification.

The model also focused on three salient regions in

the Mel-spectrogram for PD patients in the Neurovoz

dataset. These will highlight the spectrogram parts

that the model focused on because of speech features

typical for a PD, such as reduced tempo and

frequency modulation. More regions suggest that

Neurovoz contains more subtle or dispersed

Parkinsonian features, and the model needs to

consider more significant parts of the spectrograms to

make a more confident classification.

Figure 6: LIME explanations for both PD and HC from

Neurovoz dataset.

Comparing the results from the PC-GITA with the

Neurovoz datasets, it is apparent that this model

requires fewer regions in the dataset on which it was

trained to differentiate between PD and HC patients.

This fact could support the claim that the acoustic

features are more salient in the PC-GITA dataset;

therefore, the model can rely on fewer key areas for

classification. In contrast, the intensive distribution of

important regions in the Neurovoz dataset suggests

that more dispersed and subtle features need to be

captured by the model. These differences are likely

attributable to variations in the demographics of

speakers, languages, or recording conditions between

the two datasets.

LIME has provided much more insight for the

clinician into the model's decision-making process,

and the highlighted regions in the Mel-spectrograms

point to specific vocal features relevant in a clinical

sense for distinguishing PD from HC patients. For

instance, monotonic speech, reduced articulation, or

slowed speech are well-established indicators of PD

and parameters that the model has paid much

attention to agree with the clinical expectation. With

interpretable explanations, LIME ensures the

XGBoost model provides transparent predictions that

the patient can clinically validate. This

interpretability is necessary for embedding the ML

model into clinical decision-making to enable early

detection of PD.

4.2 Comparative Analysis with

Previous Studies

Results obtained with the proposed model are

compared to several recent related works on the early

detection of PD using ML models and speech

features, focusing on the accuracy achieved in

detecting PD from the sustained phonation of the

vowel /a/ across different models and datasets

outlined in Table 4. During the last few years, various

studies have investigated the potential of ML models

in diagnosing PD through speech signal processing.

These works have extracted different acoustic

features from voice records and applied various ML

techniques to attain higher diagnostic accuracy.

Recently, Wodzinski et al. (Wodzinski et al.,

2019) conducted a serious study on the use of MFCC

features that were then converted into spectrogram

images for analysis. The method was based on a

ResNet convolutional neural network model, which

treated the problem as an image classification task.

This research utilized the PC-GITA dataset for its

study and powered an accuracy of 91.7% regarding

detecting PD on the sustained phonation of the vowel

sound /a/. They have used deep learning techniques to

show how voice data can generally be analyzed to

detect diseases.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

280

Table 4: Comparison of performance of the proposed model with previous studies.

References Input Model Dataset

Accurac

Wodzinski M. et

al.

MFCCs transforming them into

ectrogram images

ResNet PC-GITA 91.7%

Nayak S. S. et al.

MFCC + Mel-

spectrogram + Spectralcontrast +

Chroma

ram + Tonnetz with GA

SVM PC-GITA 94%

Ibarra E. J. et al. Mel-scale spectrograms 2D-CNN

eurovoz 74.9%

Our work

MFCCs with their derivatives+ STFT

spectrograms + Mel-spectrograms

transforming them into Spectrogram

images

XGBoost PC_GITA 95%

SVM Neurovoz 90.74%

Similarly, another work by Nayak et al. (Nayak et

al., 2023) featured a much higher dimensionality

feature set upon combining MFCC, Mel-

spectrogram, Spectral Contrast, Chromagram, and

Tonnetz features. These features were optimized by

GA so that the features traits of voice signals would

not go unnoticed. For this ML model, SVM was

applied with an accuracy as high as 94% on the PC-

GITA dataset. The current study underlined the

effectiveness of conventional ML models once

combined with optimized feature selection methods.

Meanwhile, Ibarra et al. (Ibarra et al., 2023)

applied deep learning by using Mel-spectrograms as

the input to the 2D CNN model. Their research was

done with the Neurovoz dataset, which has voice

recordings in many different languages; thus, it is

more diverse than PC-GITA on the one hand. The

model, however, performed worse, with a

performance accuracy of 74.9%. This decrease in the

accuracy could be due to difficulties the model has

faced with generalizing due to the various languages

and features of the voices. Nevertheless, the presented

study gave insight into the perspective of performing

a language-independent PD screening.

This work adopted a representation combining

MFCCs with their derivatives, STFT spectrograms,

and Mel-spectrograms. Adopting the SMOTE

technique to overcome the data imbalance and

GridSearch for hyperparameter tuning resulted in the

proposed XGBoost, trained on PC-GITA, reaching an

accuracy of 95%, outperforming the state-of-the-art

ML algorithms. On the other hand, our SVM model,

trained on Neurovoz, yielded an accuracy of 90.74%,

proving the benefits of using intense feature

extraction and balancing techniques.

Our approach tends to outperform other types in

the proposed PC-GITA dataset, showing the

effectiveness of incorporating ML with rich acoustic

features and a strong preprocessing technique. This

comparison underlines our approach's potential for

enhancing early PD detection by voice analysis and

ML.

5 CONCLUSION

The examination performed in this paper outlines the

capability of ML algorithms focused on SVM and

XGBoost for early detection of PD by analyzing

spectrograms extracted from speech recordings.

Feature extraction techniques were enhanced based

on MFCCs with their derivatives, STFT, and Mel-

Spectrograms combined with class balancing through

SMOTE and hyperparameter tuning via

GridSearchCV, allowing our models to produce

promising results for two datasets. XGBoost was the

top classifier with an AUC of 95% in the PC-GITA

dataset. On no account did SVM fail to prove its

worth in both datasets and qualify as one of the ideal

candidates for clinical usage. By applying LIME, the

interpretability of the models was advanced and

provided helpful insight into the most predictive

vocal features, catering to a significant requirement

for the model

's clinical deployment. Nonetheless,

variations in model performance across different

datasets outline the necessity of optimization for

particular datasets and further exploration of feature

extraction techniques. Subsequent efforts should

prioritize enhancing dataset variety, refining models,

and including more acoustic cues to augment

generalization and the precision of early PD detection

systems. These discoveries will facilitate the

development of non-invasive diagnostic instruments

that allow for early identification and prompt

intervention in PD.

A Comparative and Explainable Study of Machine Learning Models for Early Detection of Parkinson’s Disease Using Spectrograms

281

DATA AVAIBILITY STATEMENT

The PC-GITA and Neurovoz datasets are available

from the authors upon request.

REFERENCES

Badhan, P. K., & Kaur, M. (2024). Early Detection of

Parkinson Disease Throughbiomedical Speech and

Voice Analysis. International Journal on Soft

Computing, Artificial Intelligence and Applications,

13(1), 11–22. https://doi.org/10.5121/ijscai.2024.13102

Bind, S., Tiwari, A. K., & Sahani, A. K. (2015). A Survey of

Machine Learning Based Approaches for Parkinson

Disease Prediction. 6.

Bloem, B. R., Okun, M. S., & Klein, C. (2021). Parkinson’s

disease. The Lancet, 397(10291), 2284–2303.

https://doi.org/10.1016/S0140-6736(21)00218-X

Breiman, L. (2001). Random Forests. Machine Learning,

45(1), 5–32. https://doi.org/10.1023/A:1010933404324

Brownlee, J. (2020). Imbalanced Classification with

Python: Better Metrics, Balance Skewed Classes, Cost-

Sensitive Learning. Machine Learning Mastery.

Dorsey, E. R., Elbaz, A., Nichols, E., Abbasi, N., Abd-

Allah, F., Abdelalim, A., Adsuar, J. C., Ansha, M. G.,

Brayne, C., Choi, J.-Y. J., Collado-Mateo, D.,

Dahodwala, N., Do, H. P., Edessa, D., Endres, M.,

Fereshtehnejad, S.-M., Foreman, K. J., Gankpe, F. G.,

Gupta, R., … Murray, C. J. L. (2018). Global, regional,

and national burden of Parkinson’s disease, 1990–2016:

A systematic analysis for the Global Burden of Disease

Study 2016. The Lancet Neurology, 17(11), 939–953.

https://doi.org/10.1016/S1474-4422(18)30295-3

ffmpeg-normalize: Normalize audio via ffmpeg (Version

1.28.3). (n.d.). [Python]. Retrieved August 28, 2024,

from https://github.com/slhck/ffmpeg-normalize

Gullapalli, A. S., & Mittal, V. K. (2022). Early Detection of

Parkinson’s Disease Through Speech Features and

Machine Learning: A Review. In T. Senjyu, P. N.

Mahalle, T. Perumal, & A. Joshi (Eds.), ICT with

Intelligent Applications (pp. 203–212). Springer

Singapore.

Harel, B., Cannizzaro, M., & Snyder, P. J. (2004).

Variability in fundamental frequency during speech in

prodromal and incipient Parkinson’s disease: A

longitudinal case study. Brain and Cognition, 56(1),

24–29. https://doi.org/10.1016/j.bandc.2004.05.002

Ibarra, E. J., Arias-Londoño, J. D., Zañartu, M., & Godino-

Llorente, J. I. (2023). Towards a Corpus (and

Language)-Independent Screening of Parkinson’s

Disease from Voice and Speech through Domain

Adaptation. Bioengineering, 10(11), 1316.

https://doi.org/10.3390/bioengineering10111316

Jumanto, J., Rofik, R., Sugiharti, E., Alamsyah, A.,

Arifudin, R., Prasetiyo, B., & Muslim, M. A. (2024).

Optimizing Support Vector Machine Performance for

Parkinson’s Disease Diagnosis Using GridSearchCV

and PCA-Based Feature Extraction. Journal of

Information Systems Engineering and Business

Intelligence, 10(1), 38–50. https://doi.org/10.

20473/jisebi.10.1.38-50

Klempíř, O., & Krupička, R. (2024). Analyzing Wav2Vec

1.0 Embeddings for Cross-Database Parkinson’s

Disease Detection and Speech Features Extraction.

Sensors, 24(17), Article 17. https://doi.org/10.

3390/s24175520

Mendes-Laureano, J., Gómez-García, J. A., Guerrero-

López, A., Luque-Buzo, E., Arias-Londoño, J. D.,

Grandas-Pérez, F. J., & Godino-Llorente, J. I. (2024).

NeuroVoz: A Castillian Spanish corpus of parkinsonian

speech (arXiv:2403.02371). arXiv. http://arxiv.

org/abs/2403.02371

Mishra, S. P., Warule, P., & Deb, S. (2024). Speech emotion

recognition using MFCC-based entropy feature. Signal,

Image and Video Processing, 18(1), 153–161.

https://doi.org/10.1007/s11760-023-02716-7

Molnar, C. (2020). Interpretable Machine Learning.

Lulu.com.

Murman, D. (2012). Early treatment of Parkinson’s disease:

Opportunities for managed care. The American Journal

of Managed Care, 18, s183-8.

Nayak, S. S., Darji, A. D., & Shah, P. K. (2023).

Identification of Parkinson’s disease from speech signal

using machine learning approach. International

Journal of Speech Technology, 26(4), 981–990.

https://doi.org/10.1007/s10772-023-10068-3

Orozco-Arroyave, J. R., & Noth, E. (n.d.). New Spanish

speech corpus database for the analysis of people

suffering from Parkinson’s disease.

Wang, X., Chen, X., Wang, Q., & Chen, G. (2022). Early

Diagnosis of Parkinson’s Disease with Speech

Pronunciation Features Based on XGBoost Model.

2022 IEEE 2nd International Conference on Software

Engineering and Artificial Intelligence (SEAI), 209–

213. https://doi.org/10.1109/SEAI55746.2022.9832191

Wodzinski, M., Skalski, A., Hemmerling, D., Orozco-

Arroyave, J. R., & Nöth, E. (2019). Deep Learning

Approach to Parkinson’s Disease Detection Using

Voice Recordings and Convolutional Neural Network

Dedicated to Image Classification. 2019 41st Annual

International Conference of the IEEE Engineering in

Medicine and Biology Society (EMBC), 717–720.

https://doi.org/10.1109/EMBC.2019.8856972

Xuan, O. Z. (2023, April 24). Exploring the Short-Time

Fourier Transform: Analyzing Time-Varying

Audio Signals. Medium. https://medium.com/

@ongzhixuan/exploring-the-short-time-fourier-

transform-analyzing-time-varying-audio-signals-

98157d1b9a12.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

282