Earlier prediction of Parkinson’s disease using cross non-decimated wavelet transform and machine learning algorithm

Abstract

Parkinson’s disease (PD) is a brain disorder, that affects a person’s body movement causing stiffness, shaking and imbalance. Earlier detection of PD is a challenging task for researchers. In this paper, earlier detection of PD is performed using the Cross-Non-Decimated Wavelet Transform (CNDWT) and Bayesian Optimized Multiple Linear Regression (BOMLR) algorithm. The PD voice data is amplitude-sliced, augmented and processed with CNDWT. The CNDWT decomposes amplitude-sliced augmented PD voice data with the Haar transform and reconstruction is performed using Daubechies wavelet of order 3 (DB3) transformations. Preprocessed data is correlated for identification of highly influential data attributes of PD. The Bayesian Optimized Multiple Linear Regression (BOMLR) method is applied to the highly correlated data attributes for earlier PD prediction. A voice data set is created in this study, which consists of 31 voice recordings of which 23 are from individuals affected by PD. The proposed CNDWT method is compared with existing methods. The results show that the proposed CNDWT method outperforms other traditional algorithms with an accuracy of 99% in predicting Parkinson’s disease.

Introduction

Parkinson’s disease (PD) is a neurodegenerative disorder that primarily affects the regions of the brain responsible for controlling body movements. Parkinson’s disease is progressive and causes motor impairments such as shakiness in the hands, changes in walking style, talking, sleeping, and thinking patterns. Patients affected by PD need support and management based on their symptoms¹. In the human brain, the substantia nigra, which is located in the basal ganglia, plays a key role in controlling body movements. This region consists of cells that produce a chemical called dopamine, which helps to transmit messages to the brain. Dopamine transmits messages to the nervous system and controls the body’s movements. When PD occurs, the substantia nigra cells lose the ability to carry messages due to a deficiency of dopamine levels, causing a significant impact on body movement².

In the early stages of PD, symptoms cannot be easily observed. As time progresses, an increase in the number of dead cells leads to a peak level of PD, and by the time the disease is identified, almost 80% of the cells may be dead, often without realizing the cause. Motor symptoms of PD include tremors, rigidity, postural instability, and bradykinesia. Non-motor symptoms of PD include depression, bladder problems, and abnormal movements. Earlier diagnosis of PD improves patients’ health through appropriate treatment. Initially, vocal impairments in PD patients occur at an earlier stage of the disease³. One of the major causes of Parkinson’s disease is Normal Pressure Hydrocephalus (NPH), which influences the lower part of the body⁴. The symptoms include difficulty in walking, excessive urination, and memory-related issues. Aging is considered the primary risk factor for PD, as it is associated with a natural decline in cellular function and regenerative capacity⁵. By the year 2050, the prevalence of Parkinson’s disease is projected to reach 267 cases per 100,000 population (ranging from 230 to 320), representing a substantial increase of 76% (56% to 125%) compared to 2021. Similarly, the age-standardized prevalence is expected to rise to 216 per 100,000 (with a range of 168 to 281), marking a 55% increase (50% to 60%) from 2021 levels⁶.

There are multiple methods for detecting and predicting Parkinson’s disease, such as handwriting analysis, voice analysis, electroencephalography (EEG), magnetic resonance imaging (MRI), and functional MRI (fMRI). However, EEG, MRI, and fMRI require clinical diagnoses, which involve specific laboratory tests using advanced equipment. Patients often face long wait times for these examinations, and many individuals may have limited access to such facilities. The degeneration of dopaminergic neurons can occur rapidly; by the time a clinical diagnosis is made, significant neuron loss may have already occurred. In contrast, the most accessible tests are voice and handwriting analyses⁷. This paper focuses on using voice data from patients to predict Parkinson’s disease through the analysis of sustained phonation and speech patterns.

Figure 1 shows the detection modalities of Parkinson’s disease, which include voice analysis of PD-affected patients and neuroimaging techniques such as MRI and fMRI that visualize the brain structures and their activities related to PD symptoms. Dopamine transporter scans assess the dopamine levels in the brain, and finally, clinical evaluations of the motor and non-motor symptoms associated with PD.

Fig. 1.

PD detection modalities.

To predict PD at an early stage, this article proposes a Cross Non-Decimated Wavelet Transform model for preprocessing the input voice data of individuals with PD. The dataset used for this study is the PD voice dataset from Kaggle, which consists of voice recordings from 31 people, of which 23 are PD patients and 8 are healthy individuals. In the proposed CNDWT method, amplitude slicing is applied to the raw data to isolate and emphasize subtle variations in speech caused by vocal impairments. This process helps reduce noise and generate additional samples, thereby enhancing the model’s performance. Then, two wavelet transforms, namely cross DWT and cross NDWT, are applied to the preprocessed signals for decomposition and reconstruction. DWT decomposes the signal into different frequency sub-bands, making it easier to analyze temporal and spectral variations, such as tremors or irregularities in speech. However, DWT involves downsampling, which makes it shift-variant—a drawback for time-sensitive biomedical signals. On the other hand, NDWT (also called stationary wavelet transform) avoids downsampling, preserving the alignment of signal features across scales and ensuring shift-invariance, which is essential for accurately reconstructing and analyzing voice patterns affected by PD. This makes NDWT especially useful in medical diagnostics, where fine-grained, consistent feature representation is critical.

This study introduces a modified version of the Non-Decimated Wavelet Transform (NDWT), referred to as Cross NDWT (CNDWT), for preprocessing the Parkinson’s disease (PD) voice dataset. In CNDWT, the Haar wavelet transform is applied for signal decomposition, while the Daubechies-3 (DB3) wavelet is used for signal reconstruction. Following preprocessing, a correlation analysis is conducted to identify relationships among the extracted attributes. Understanding these correlations is crucial for feature relevance, enabling effective feature selection, handling of missing data, reduction of overfitting, and ultimately enhancing prediction accuracy. The highly correlated features are then used for PD prediction using Bayesian Optimized Multiple Linear Regression (BOMLR). Finally, these highly correlated attributes are used for the prediction of PD using Bayesian Optimized Multiple Linear Regression (BOMLR).

Problem statement

Parkinson’s disease (PD) is a progressive neurological disorder that affects motor and speech functions. Traditional diagnostic approaches such as MRI or EEG, while effective, are costly, require clinical supervision, and are not easily accessible to all patients. Voice analysis has emerged as a non-invasive and accessible method for early detection of PD², but existing methods often fail to capture subtle vocal impairments due to noise, limited feature selection, and lack of optimized predictive modeling. There is a pressing need for a robust preprocessing and classification framework that enhances voice signal quality and accurately predicts PD in its early stages. This study addresses this gap by proposing a novel Cross NDWT-based preprocessing approach combined with Bayesian Optimized Multiple Linear Regression for improved PD prediction using voice data.

Contributions

The contributions of the proposed CNDWT method for earlier detection of PD are as follows:

• i.To develop an amplitude slicing-based data augmentation technique to enhance the representation of PD voice signals for early-stage detection.
• ii.To propose a novel preprocessing framework, Cross-NDWT, which employs the Haar wavelet for signal decomposition and denoising, and the DB3 wavelet for signal reconstruction.
• iii.To identify highly correlated voice attributes from the augmented PD dataset that are most indicative of early Parkinson’s symptoms.
• iv.To implement a Bayesian Optimized Multiple Linear Regression (BOMLR) model to predict PD based on the selected correlated features.
• v.To evaluate and compare the performance of the proposed CNDWT-BOMLR framework against traditional preprocessing and classification methods, focusing on precision and accuracy for early PD detection.

The remaining content of this article is organized as follows: Sect. “Literature review” lists the literature review conducted regarding this article, Sect. “Methodology” explains the methodology of the proposed CNDWT method, Sect. “Results” displays the results and discussions, and finally, Sect. “Discussion” presents the conclusion.

Literature review

This section elaborates on the work done by various researchers in predicting Parkinson’s disease using different machine learning and deep learning algorithms. Deep learning methods and machine learning methods are used for the early prediction of PD (Wang et al.⁸. The authors used the Parkinson’s Progression Markers Initiative (PPMI) database and employed biological sampling, imaging, and behavioral analysis to determine the biomarkers of PD. The authors state that the deep learning methods perform better than the machine learning models. Lamba et al.⁹ suggested three feature selection methods and three classifiers for the prediction of PD at an early stage. The authors used the UCI repository speech dataset and applied the Synthetic Minority Oversampling Technique (SMOTE) to eliminate the imbalances in the data. Genetic algorithms and random forest classifiers perform better than other methods. Salmanpour et al.¹⁰ used Principal Component Analysis (PCA) and Support Vector Regression (SVR) for the prediction of PD. The Feature Subset Selector Algorithms (FSSA) were used for feature selection. The authors used data from 204 PD patients with 18 features for prediction. Krishna et al.¹¹used the Logistic Decision Regression (LDR) algorithm along with machine learning for the prediction of Parkinson’s disease. The above methods classify PD and detect PD at an early stage. Feature selection was performed using the Least Absolute Shrinkage and Feature Selection (LAFS) algorithm. Taleb et al.¹²used a handwriting, dataset that includes handwriting samples from 42 people. The authors combined Convolutional Neural Networks (CNN) and Binary Long Short Term Memory (BLSTM) for the prediction of Parkinson’s disease using time series classification. The authors used a handwriting dataset comprising 42 subjects, 21 of whom are PD patients and 21 are healthy individuals. Techniques such as spectrograms were used to encode the pen-drawn signals into images. Shahid et al.¹³ suggested a deep neural network model to predict PD at an early stage. The Parkinson’s telemonitoring dataset was used for the prediction and assessment of the PD rating using the UPDRS. PCA was deployed for feature selection, and DNN was used for the prediction. Solana et al.¹⁴ designed a PD detection algorithm with vocal features. The reduction in the subset was performed using the Wrapper subset selection method, and the following four classifiers were used for PD prediction: K-Nearest Neighbor, SVM, Random Forest, and Multi-Layer Perceptron. Bind et al.¹⁵, reviewed the machine learning methods for the prediction of Parkinson’s disease. ML methods such as support vector machines and Naïve Bayes were compared, and the results of previous works that used these methods were listed and analyzed. Parkinson’s disease is predicted by a person’s smell signatures, using VOC sensors to identify the smell through sweat in the person (Kulkarni et al.¹⁶. Arduino UNO was used as an interface for sensing the smell. PD detection was performed using speech signals by Karan et al.¹⁷. The author found that PD alters speech intelligibility through components such as intensity, respiration, phonation, and articulation. Empirical Mode Decomposition (EMD) was applied to the speech signal, which decomposes into Intrinsic Mode Functions (IMF). The IMF of vocal tract and vocal fold vibration information determines the non-linear characteristics of the speech signal. Table 1 shows some of the existing methods and the data types used for the prediction of PD. Table 2 shows the statistics of the previous studies carried out and the type of input used for predicting PD (www.worldcat.com).

Table 1.

Survey of datasets and methods used in existing articles.

Articles	Dataset	Methods	Accuracy (%)
[9]	Speech	Genetic Algorithm/Random Forest	95
[11]	Handwriting	CNN-BLSTM	97
[14]	Speech
[18]	EEG	TQTWT	96
[19]	EEG	Automated TQTWT	96
[20]	Non-Motor	Wrapper feature selection	93
[21]	MRI	Reflexive saccades measurements	70
[22]	MRI	SVM	70
[23]	Recorded Speech Signal	Random Tree Classification	85
[24]	Dyskinesia Data	SVM	77

Table 2.

Statistics of existing work done in PD prediction.

Type of dataset	Articles	Downloadable article	Print books	E Books
Voice	250	15	9	31
Handwriting	76	3	2	12
EEG	740	56	59	160
MRI	1200	210	53	130

Methodology

The dataset is augmented using amplitude slicing and is used for PD detection at an earlier stage. The voice data collected from PD patients was only after a prolonged period of PD symptoms. To study the initial stage of PD, the dataset is amplitude sliced, which mimics the PD condition at an early stage. Figure 2 shows the proposed CNDWT method for PD detection at an earlier stage. The amplitude sliced augmented PD voice dataset is given as input for preprocessing. The raw dataset cannot be used directly for PD detection. Preprocessing is required for the extraction of feature variables for better classification²⁵. The proposed Cross NDWT (CNDWT) is used for signal decomposition and noise reduction. CNDWT performs better than DWT. Correlation is applied to find the relationship between the feature variables after preprocessing. Once the correlation between the feature variables is identified, only highly correlated attributes that significantly influence PD are selected. Nine attributes are highly correlated and have a strong influence on PD. Using these highly influential attributes, the BOMLR method is applied to predict Parkinson’s disease at an earlier stage.

Fig. 2.

Work Flow of the proposed CNDWT method.

Figure 3 shows the flow chart of the proposed CNDWT method, where the PD voice dataset is given as input for preprocessing using the Haar NDWT, and reconstruction of the signal is performed using DB3 NDWT. After preprocessing, the correlation among the attributes is computed, and highly correlated attributes are selected for PD prediction using the Bayesian Optimized Multiple Linear Regression Model (BOMLR).

Fig. 3.

Flow Chart the proposed CNDWT Method.

PD voice dataset

The proposed CNDWT method uses the PD voice dataset obtained from Kaggle, which consists of voice recordings from 31 individuals, including 23 patients diagnosed with Parkinson’s disease (PD) and 8 healthy controls. The dataset is publicly available at https://www.kaggle.com/datasets/vikasukani/parkinsons-disease-data-set. The dataset description is given in Table 3. Each participant recorded multiple samples, such as sustained vowels, words, phrases, and numerals. Various acoustic vocal features relevant for differentiating PD patients from healthy individuals were extracted from the voice recordings. Each row in the dataset corresponds to a single voice recording sample. The features include a wide range of time- and frequency-domain parameters such as jitter, shimmer, HNR (Harmonic-to-Noise Ratio), MFCCs (Mel-frequency cepstral coefficients), and more. A binary label column named “status” is provided, where 1 indicates a PD patient and 0 indicates a healthy control. The recordings were originally collected and curated by clinical researchers as part of a medical study. The dataset was inspected for any missing or null values, and there were no missing columns. No additional exclusion criteria were applied beyond what was defined in the original dataset. All available samples were used in the analysis. To increase the diversity of training data and mitigate the imbalance between healthy and PD samples, the proposed method employed amplitude slicing augmentation. This method segments the signal amplitude into slices, producing variations of existing samples while preserving pathological vocal characteristics. Table 4 gives the reason for choosing the voice dataset for PD prediction, and Table 5 lists out the comparison of the voice data and other modalities that are used for PD prediction.

Table 3.

Dataset Description.

Description	Values
Dataset	PD Voice Dataset - Kaggle
Data Preprocessing	Amplitude Slicing Augmentation
PD Voice Data – Before Augmentation	Total Subjects − 31 Parkinson Subjects − 23 Healthy Subjects − 8
Synthetic Data Added for Augmentation	Total Subjects − 630 Parkinson Subjects – 300 Healthy Subjects − 330
PD Voice Data - After Augmentation	Total Subjects − 661 Total Parkinson Subjects – 323 Total Healthy Subjects – 338

Table 4.

Reason for choosing voice data as input.

Aspects	Reason	Description
Early Symptom Manifestation	Vocal impairments often appear in the early stages of PD.	Allows for earlier diagnosis compared to motor or cognitive symptoms.
Non-Invasive Data Collection	Voice samples can be collected without physical contact or clinical intervention.	Makes the process safer, patient-friendly, and scalable.
Accessibility and Cost	Recording voice requires minimal infrastructure.	Suitable for remote, low-resource, or large-scale screening.
Frequency and Intensity Changes	PD causes monotonic speech, reduced vocal loudness, and voice tremors.	These changes are quantifiable and indicative of PD progression.
Compatibility with AI Models	Voice features (e.g., jitter, shimmer, HNR) are well-suited for machine learning models.	Facilitates high-accuracy prediction through feature engineering and deep learning techniques.
Avoids Expensive Diagnostics	Alternatives like MRI or EEG are expensive and not widely accessible.	Voice analysis provides a cost-effective substitute for initial screening.

Table 5.

Comparison of voice data and other modalities for PD prediction.

Modality	Cost	Clinical Setup	Data Collection Location	Ease of Collection	Availability	Capability of Early Prediction
Voice	Low	No	Home	Easy	Widely Available	Yes
Handwriting	Moderate	No	Home	Moderate	Moderately Available	Moderate
EEG	High	Yes	Hospital/Clinic	Difficult	Limited	No
MRI/fMRI	Very High	Yes	Hospital/Clinic	Very Difficult	Highly Limited	No

Amplitude slicing augmentation

Data augmentation is employed in various classification processes to overcome the problems of overfitting and data imbalance. Applications such as face detection, medical imaging, and voice detection utilize augmentation techniques, especially when working with small datasets. Signal processing applications often require methods such as scaling, rotation, shifting, and the addition of noise. In this paper, amplitude slicing augmentation is proposed to create variations of the original voice recordings and generate synthetic data to address data imbalance.

Amplitude slicing augmentation is a data augmentation technique applied to voice signals, wherein the amplitude envelope is segmented into discrete levels or “slices.” These slices may be selectively modified or replaced to simulate variations in signal intensity while preserving the temporal and spectral characteristics of the original waveform. This process introduces controlled amplitude variability in the training data, effectively mimicking real-world conditions such as differences in speaker loudness, recording equipment, or acoustic environments²⁶. The application of amplitude slicing augmentation impacts the input voice data by altering its energy profile without affecting critical attributes such as pitch, timing, or phonetic content. As a result, this technique enhances the robustness and generalization capability of voice-based models by exposing them to a broader range of amplitude conditions during training²⁷. By applying amplitude slicing augmentation to the input voice data, new synthetic data are created to avoid data imbalance in the original PD voice data.

The input PD voice data are amplitude slicing augmented, which mimics the early stage of PD voices. The transformations that are applied to the original voice data to obtain synthetic data are listed in Table 6 Normalization standardizes the amplitude across all audio samples to ensure uniformity and reduce variability from different recording conditions. Segmentation then divides the signal into slices based on amplitude levels, enabling focused analysis of specific signal regions. From each of these slices, feature extraction is performed to capture detailed and relevant characteristics that contribute to model learning. Finally, augmentation uses these processed slices to generate new or enhanced data samples, thereby improving the diversity and quality of the dataset for more robust model training.

Table 6.

Transformations applied to the voice data using amplitude slicing for augmentation.

Transformation	Purpose/Effect
Normalization	Standardizes amplitude across samples
Segmentation	Divides signal into amplitude-based slices
Feature Extraction	Captures detailed info from each slice
Augmentation	Generates new samples or enhances data quality

The original Parkinson’s disease (PD) voice dataset obtained from Kaggle is imbalanced, with a higher number of samples from Parkinson’s subjects compared to healthy controls. To address this imbalance and reduce the risk of overfitting, new data samples are generated using amplitude slicing augmentation. This approach helps balance the number of Parkinson’s and healthy subject samples in the dataset. A total of 31 subjects were included in the original Kaggle dataset, comprising 23 individuals affected by Parkinson’s disease (PD) and 8 healthy controls. After data augmentation, 630 samples of synthetic data were generated, with 300 samples representing PD-affected subjects and 330 representing healthy subjects. The complete training dataset consists of 661 samples, including 323 Parkinson’s subjects and 338 healthy individuals. The information regarding the acoustic vocal feature attributes is listed in Table 7, and Fig. 4 shows sample data from the PD voice dataset.

Table 7.

Vocal feature attributes.

S.No	Abbreviations	Feature description
1.	MDVP: F0 (Hz)	Average vocal fundamental frequency
2.	MDVP: Fhi (Hz)	Maximum vocal fundamental frequency
3.	MDVP: Flo (Hz)	Minimum vocal fundamental frequency
4.	MDVP: Jitter(%)	MDVP jitter in percentage
5.	MDVP: Jitter(Abs)	MDVP absolute jitter in ms
6.	MDVP: RAP	MDVP relative amplitude perturbation
7.	MDVP: PPQ	MDVP five-point period perturbation quotient
8.	Jitter: DDP	Average absolute difference of differences between jitter cycles
9.	MDVP: Shimmer	MDVP local shimmer
10.	MDVP: Shimmer(dB)	MDVP local shimmer in dB

Fig. 4.

Sample data from PD voice data set.

Model selection and techniques implemented

The choice of algorithms and techniques in this study was guided by the nature of the dataset and the goal of accurately distinguishing between Parkinson’s Disease (PD) patients and healthy controls based on acoustic voice features.

• Classification Algorithms: Multiple classification algorithms were evaluated, including Support Vector Machines (SVM), Random Forest (RF), and Logistic Regression, as these are well-established in biomedical signal classification tasks and perform well on structured data with moderate dimensionality. Among them, the proposed Multiple Linear Regression (MLR) model was selected as the primary classifier due to its robustness in handling high-dimensional data and its ability to create complex decision boundaries that can effectively separate overlapping classes.

• Feature Set: The dataset includes clinically relevant acoustic features such as jitter, shimmer, HNR, and MFCCs, which are known indicators of voice disorders associated with Parkinson’s disease. These features were used directly without dimensionality reduction to retain all available discriminative information.

• Data Augmentation: Amplitude slicing augmentation was employed to increase sample diversity and address class imbalance. This technique was chosen because it generates meaningful variations in the signal amplitude while preserving critical pathological characteristics, helping prevent overfitting and improving model generalization.

• Parameter Tuning: Hyperparameters for each model were optimized using Bayesian optimization, ensuring fair evaluation and preventing information leakage.

• Evaluation Strategy: A leave-one-subject-out (LOSO) cross-validation approach was used to ensure that the class distribution was maintained across training and validation folds. This provided a more reliable estimate of model performance, especially given the class imbalance.

Cross-discrete wavelet transform (C-DWT)

DWT is a signal processing technique used for denoising and decomposing the raw input signal into a set of wavelet coefficients or functions using one single function. These functions are of the same length, and DWT truncates the data to extract the important coefficients, while the other coefficients are replaced by 0. This process reduces the size of the data and eliminates unimportant information, which makes DWT a robust classification model compared to other transformations. DWT is well suited for time-series data and is therefore used in predicting PD, as a large amount of data is required for prediction, and data compression becomes necessary. Reconstruction of the signal is performed by inverting the DWT for the specified set of coefficients²⁸. In DWT, the data is transformed into wavelet vectors of the same length. DWT truncates the transformed data and extracts the most relevant wavelet coefficients, which reduces the dimensions of the data by altering the other coefficients to zero. These reductions make DWT a reliable multidimensional model for classification compared to other wavelet transforms. It is an efficient technique for time series data-based applications. DWT is more suitable for monitoring the time series data of PD-affected persons, which are usually huge in size, and data reduction becomes mandatory. In this study, Haar DWT is used for signal decomposition, and db3 DWT is applied for signal reconstruction and is termed C-DWT.

Cross-non-decimated wavelet transform (C-NDWT)

The proposed CNDWT method uses Non-Decimated Wavelet Transform (CNDWT) for processing voice signals. This wavelet transform never uses downsampling and is addressed as shift invariant, which is suitable for stationary or non-stationary signals. A comparison between DWT and NDWT based on their performance was carried out using real-time and simulation applications. The highlight of non-decimated wavelet transforms is the ability to be shift invariant, restoring the accurate signal and aiding in signal pattern detection. DWT and NDWT fall under discrete wavelet transforms. The two wavelet transforms vary in their scale and location parameters. The continuous wavelet transform allows for continuous or discrete values, whereas the discrete wavelet transform utilizes discrete values, often powers of two. In the case of the NDWT, the outputs do not undergo downsampling, resulting in scales possessing identical wavelet coefficients²⁹. The discrete wavelet transform uses filters with a limited number of coefficients, ensuring efficient implementation. NDWT’s shift-invariant nature and consistent number of coefficients at all scales make it more suitable for signal processing. This enables the detection of hidden features in stationary and non-stationary signals in time and frequency locations.

HAAR and daubechies wavelet (DB3) for PD speech signal (C-NDWT)

The orthogonal wavelet is the Haar function, which is used for translation or shifting and dilation or scaling. This Haar wavelet removes the high-frequency components and smoothens the signal. It decomposes the original signal into two sub-signals, which are equal in length. The coefficients of the Haar wavelet develop the following four sub-bands: LL, LH, HL, and HH. During the first level of decomposition, low and high pass filters are used, yielding approximation and detail coefficients, namely (A1) and (D1) respectively. This approximation (A1) is decomposed recursively, and the termination is computed based on the signal’s frequency components. In this article, the Haar transform decomposes the signal, and the DB3 wavelet transformation is used for reconstruction. Initially, Haar will decompose every processed signal from the channel into an approximate coefficient (A1) and four more detail coefficients D1 through D4. These coefficients have frequency values in multiples of 16, ranging from 0 to 256, and each of them possesses 16 Hz of sub-band. In wavelet-based time–frequency analysis, there is an inherent trade-off: higher-frequency components are analyzed with better time resolution but poorer frequency resolution, while lower-frequency components receive better frequency resolution at the expense of time precision. This behavior stems from the scaling nature of the wavelet transform, which effectively uses shorter time windows for high frequencies and longer windows for low frequencies, thus adapting resolution dynamically³⁰. During the first level of decomposition, low and high pass filters are used, yielding approximation and detail coefficients, namely (A1) and (D1) respectively³¹. This approximation (A1) is decomposed recursively, and the termination is computed based on the signal’s frequency components.

Daubechies wavelet belongs to the family of orthogonal wavelets used in signal preprocessing and compression. The number 3 in “DB3” indicates the number of vanishing moments in the wavelet transform. This type of wavelet has a better balance between frequency and time localization properties. Daubechies wavelets have overlapping windows; therefore, the higher frequency spectrum reflects entire high-frequency changes, and hence they are well suited for noise removal of PD signals³². Table 8 shows the difference between DWT and CNDWT.

Table 8.

Difference between DWT and CNDWT.

Aspect	DWT	CNDWT
Time-Frequency resolution	Fixed resolution; loses phase info due to decimation.	Higher resolution with non-decimated sub bands; retains phase info for vocal nuances.
Shift invariance	Sensitive to signal shifts (decimation causes aliasing).	Shift-invariant (no decimation), better for vocal variability.
Phase information	Discards phase, critical for voice harmonics and dysphonia detection.	Captures phase, improving feature discriminability in PD voice signals.
Noise robustness	Vulnerable to high-frequency noise in detail coefficients.	Reduced noise sensitivity through complex magnitude averaging.
PD detection impact	Less sensitive to vocal impairments, reduced classification performance	More sensitive to vocal features typical in PD (e.g., jitter, shimmer), boosting classification

Correlation analysis of attributes

Correlation quantifies the strength and direction of the relationship between two variables. For instance, if variables x and y are positively correlated, an increase in x corresponds to an increase in y, indicating a proportional relationship. If changes in one variable do not affect the other, there is no correlation. A weak correlation occurs when changes in one variable only slightly affect the other. Correlation can be either positive or negative: a positive correlation indicates that both variables increase or decrease together, while a negative correlation means that an increase in one leads to a decrease in the other³³. Understanding correlation is essential, as it reflects the degree of interdependence between variables. Graphically, a linear correlation appears as a straight line, while a non-linear correlation forms a curve. In machine learning, extensive time is devoted to data preprocessing and feature selection, a process where only the most relevant features are retained to enhance model accuracy and eliminate redundant or irrelevant data. This step is crucial, as irrelevant features can degrade model performance.

In the context of Parkinson’s disease (PD) detection, many vocal features display overlapping and non-linear behaviors. For example, features such as Shimmer: APQ3, MDVP: APQ11, and Shimmer: APQ5 all relate to amplitude variation. Thus, analyzing correlations among vocal features is necessary to reduce redundancy. The proposed CNDWT method calculates pairwise correlations between features in the voice dataset and selects those with high correlation coefficients (greater than 0.8) for prediction. Based on this analysis, the following highly correlated features were chosen as inputs for the machine learning model: (i) MDVP: Jitter (%), (ii) MDVP: Jitter (Abs), (iii) MDVP: RAP, (iv) MDVP: PPQ, (v) Jitter: DDP, and (vi) NHR, as shown in Table 9.

Table 9.

Analysis of parkinson’s disease vocal Attributes.

Voice Data Attributes					Correlation %
HNR The amount of excess noise in the given voice signal	MDVP: Shimmer (dB) Differences among successive amplitude	MDVP: Fhi(Hz) analyses voice patterns to detect dysphonia, a common symptom of Parkinson’s disease	MDVP: Flo(Hz)	STATUS Healthy-0 Parkinson’s-1	97%
MDVP: Shimmer(dB) Differences among successive amplitude	HNR The amount of excess noise in the given voice signal	PPE Indicates abnormal control in the pitch of the speech signal	DFA Measurement of the Self-Affinity of the voice input	STATUS Healthy-0 Parkinson’s-1	96%
Shimmer: APQ5 Indicates the perturbation ratio of the amplitude of the signal	MDVP: APQ Represents the differences in speech signal’s vibration and voice abnormality	Shimmer: DDA Higher value of this feature shows abnormality in the voice input	NHR Ratio between the amount of noise and the harmonics of the voice signal	STATUS Healthy-0 Parkinson’s-1	24%
Jitter: DDP greater variation in fundamental frequency	MDVP: PPQ Quantifies the pitch in the voice signal	MDVP: Shimmer, Shimmer measure the cycle-cycle variations in the amplitude of the voice input	HNR HNR denotes the ratio of noise to tonal components in the voice	STATUS Healthy-0 Parkinson’s-1	5%

Identification of influencing attributes of PD detector

After implementing CNDWT, the redundant information is cleared out, and the highly influencing attributes that cause PD are selected. Choosing these attributes results in less computation time and increases accuracy. In this paper, the highly influential attributes are MDVP: Jitter (%), MDVP: Jitter (Abs), MDVP: RAP, MDVP: PPQ, Jitter: DDP, and NHR.

Bayesian optimized multiple regression model

Multiple regression for PD classification

Multiple Regression is the degree of relationship that exists between three or more variables. It shows the relation between one of the variables, called the dependent variable, along with the combined effect of the other two variables, called the independent variables. The measure of this relationship is defined as a correlation between three or more variables or a multiple regression. Regression is used to find the relation between two variables and estimates, inserts, or infers. It analyzes the relationship between the independent and dependent variables of a given dataset and computes a regression model. Once the relationships between variables are identified, regression analysis is used by researchers to make predictions and gain valuable insights into influencing factors. Multiple Linear Regression (MLR) is based on a regression model that incorporates more than one independent variable and predicts the changes in the dependent variable³⁴. The estimation of the influence of each independent variable over the dependent variable is done using this model. The MLR model generates regression coefficients, which represent the strength of their impact on the dependent variable. For the prediction process, MLR combines multiple variables and yields the optimum prediction results of the dependent variable. The equation (f) of multiple regression of the variable M on N1, N2, N3,. NI is written as follows: . In this equation, a0 indicates the intercept, and a_1,a₂, a₃ indicate the slope in the linear equation. The efficiency of an MR model is identified by the R² value, which always lies between 0 and 1. If the R² is around 0.8 or 0.9, then the model has good prediction. Table 10 shows the advantages of the MLR method, and Table 11 gives a comparative analysis between MLR and other regression models for PD prediction.

Table 10.

Comparison of MLR with other regression Models.

Sl.No	Factor	Description
1.	Interpretability	MLR provides clear interpretability of feature importance and relationships between voice features and PD diagnosis.
2	Correlation	MLR works effectively with correlated inputs, allowing us to directly incorporate the proposed method’s correlation findings.
3	Computational Efficiency	The proposed method’s preprocessing pipeline involves complex wavelet transforms (Cross-DWT and NDWT), and MLR gives higher accuracy with lower computational demands.

Table 11.

Comparison of MLR with Other Regression Models

Regression models	Factor
	Interpretability	Complexity	Cost	Suitability for PD detection	Overfitting risk	Training time
LR	High	Low	Low	Yes	Low	Fast
SVR	Moderate	High	Moderate	Yes	High	Slow
DT	Moderate	Moderate	Low	Yes	High	Fast
RF	Low	High	Moderate	Yes	Low	Slow
MLR	High	Low	Low	Yes	Low	Fast

Results

This section evaluates the performance of the proposed CNDWT method by comparing the results of various experiments based on performance metrics such as accuracy, sensitivity, and specificity with the results of other methods for predicting Parkinson’s disease. The correlation coefficients between the highly influenced parameters are listed in Table 12. The graphs in Table 13. display the output of the correlation between the attributes.

Table 12.

Highly correlated vocal features Attributes.

Features	MDVP: Jitter (%)	MDVP: Jitter(Abs)	MDVP: RAP	MDVP: PPQ	Jitter: DDP	NHR
MDVP: Jitter (%)	1	0.936	0.9023	0.9467	0.9023	0.907
MDVP: Jitter (Abs)	0.9264	1	0.9589	0.8235	0.9589	0.897
MDVP: RAP	0.9812	0.9589	1	0.9589	1	0.9481
MDVP: PPQ	0.9743	0.8978	0.9573	1	0.9573	0.8446
Jitter: DDP	0.9812	0.9589	1	0.9373	1	0.9831
NHR	0.987	0.891	0.9207	0.8845	0.9207	1

Table 13.

Correlation between attributes of the PD Dataset

Figure 5 shows the performance of the proposed C-NDWT methods and compares them with various performance metrics. The results are depicted in graphs. In graphs (a) and (b), the Matthews Correlation Coefficient (MCC), Receiver Operating Characteristic (ROC) curve values, accuracy, and precision values of Multiple Regression are compared with traditional algorithms³⁵ suchas Random Forest, Support Vector Machine (SVM), Multi-Layer Perceptron, and Decision Trees. These results clearly show that the BOMLR model gives a better result. In graph (c), the Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R² values are compared. The BOMLR algorithm performs better than the traditional Neural Network (NN), Adaptive Neuro-Fuzzy Inference System (ANFIS), Support Vector Regression (SVR), and Self-Organizing Map Nonlinear Iterative Partial Least Squares Incremental Support Vector (SOM-NIPALS-ISVR) methods³⁶. Figure 5 (D) shows the comparison of accuracy, sensitivity, specificity, AUC, and F1 values of the proposed C-NDWT method compared with traditional machine learning algorithms³⁷. Similarly, in (e) and (f), the proposed CNDWT (Cross Transform – Haar X DB3) is compared with that of Fast Fourier Transform (FFT) and Discrete Wavelet Transform (DWT) based on performance metrics such as False Negative Rate (FNR), False Positive Rate (FPR), False Omission Rate (FOR), Negative Predictive Value (NPV), Sensitivity, Specificity, Accuracy (ACC), and Precision. The graph shows that the proposed CNDWT method (Cross Transform – Haar X DB3) performs better than all other methods.

Fig. 5.

Performance comparison of the proposed CNDWT method with other traditional methods. (A) Accuracy and precision comparison of the proposed method with other regression methods, (B) Performance evaluation of the proposed MR based on MCC and ROC values, (C) Evaluation of the proposed MR method with other methods for MAE, RMSE and R2 values, (D) Comparison of the performance of the proposed method with existing methods, (E) Comparison of error metrics for various feature extraction methods with the proposed method, (F) Performance comparison of the proposed feature extraction methods with DWT and FFT methods.

In addition to the PD voice dataset, two more datasets were used to evaluate the prediction and classification accuracy of the proposed Cross Non-Decimated Wavelet Transform (CNDWT) method. The first dataset is the Parkinson Speech Dataset (PSD), which consists of voice recordings from 20 patients diagnosed with Parkinson’s disease (PD) and 20 healthy subjects. The second dataset is sourced from the UCI Machine Learning Repository (UCIMLR) and includes recordings from 252 individuals, comprising 188 PD patients (107 men and 81 women) aged 33 to 87 years, and 64 healthy controls (23 men and 41 women) aged 41 to 82 years. Table 14 presents the evaluation results of the proposed CNDWT method across these two additional voice datasets, PSD and UCIMLR, demonstrating that the CNDWT method performs effectively for both datasets in predicting and classifying Parkinson’s disease using voice signals. Table 15 presents a comparative analysis of Parkinson’s disease prediction performance across various studies using different datasets and machine learning models. Prior works employed datasets such as UCI, PPMI, and the MDVR-KCL, utilizing classifiers like SVM, Random Forest (RF), k-NN, and Logistic Regression (LR). Reported accuracies ranged from 69% to 89%, with models like SVM and RF consistently achieving strong overall performance across studies. The proposed method in this study, which uses the PD Voice Dataset and the CNDWT-MR technique, outperformed all referenced approaches, achieving 98% accuracy, 97.35% recall, and 98% F1-score, indicating a substantial improvement in both sensitivity and specificity. These results underscore the effectiveness of the proposed approach in capturing disease-specific vocal patterns and provide strong support for its clinical applicability in early Parkinson’s detection.

Table 14.

Evaluation results of the proposed CNDWT method over other voice datasets.

Voice dataset name	Prediction accuracy	Classification accuracy	Sensitivity	Precision	F-Score
PSD	98	96.35	0.97	0.98	0.97
UCIMLR	97.45	98	0.98	0.98	0.98
PD-VDS	99	98.5	0.99	0.98	0.99

Table 15.

Evaluation results of the proposed CNDWT method over other voice datasets.

Reference	Dataset used	Methods	Performance Metrics (%)
			Accuracy	Recall	F1	Precision	Specificity
38	King’s College London (MDVR-KCL)	SVM	88.36	85.50	87.00	87.55	87.17
		DT	76.59	80.50	71.43	73.91	77.08
		NB	82.03	84.12	80.00	81.55	82.81
		LDA	87.50	89.23	86.00	86.17	87.68
39	PPMI & UCI	LR	85.71	89.00	-	81.63	87.00
		RF	85.00	89.00	-	81.00	87.00
		SVM	81.00	82.00	-	84.00	84.00
		KNN	89.00	87.00	-	88.00	85.00
40	UCI	SVM	88.0	80.0	88.0	88.0	-
		KNN	84.0	88.5	84.0	85.5	-
		DT	86.0	94.0	87.5	90.0	-
		RF	83.0	87.5	97.0	87.0	-
41	UCI	NB	73.00	77.00	79.00	73.00	78.00
		MLP	75.00	71.00	80.00	79.00	74.00
		K-NN	69.00	75.00	70.00	66.00	68.00
		RF	80.00	81.00	86.00	85.00	83.00
Proposed	PD Voice Dataset	CNDWT-MR	98.00	97.35	98.00	97.00	98.00

For training the model, 90% of the dataset was utilized, while the remaining 10% was reserved for testing. Amplitude slicing augmentation is used to handle data imbalance. The performance metrics, including precision, recall, F-Score, and accuracy, are noted. In the PD voice dataset, there are only 8 healthy subjects compared to 23 PD-affected subjects, which contributes to the low performance of the MR algorithm, yielding 68% precision, 70% recall, 68% F-Score, and 69% accuracy. To improve classification accuracy, the imbalanced data is balanced using amplitude slicing, and the LOSO cross-validation results are then recorded again. After data augmentation, the class imbalance is addressed, resulting in improved performance with 98% precision, 97% recall, 98% F-Score, and 98% accuracy. The results are shown in Fig. 6. Figure 7 shows the per-fold accuracy results using Leave-One-Subject-Out (LOSO) validation. Each bar represents the classification accuracy for a specific fold, where one subject’s data is used for testing and the rest for training. The consistently high accuracy across all folds (around 97–98%) indicates that the model performs reliably across different subjects, demonstrating strong generalization capability for early prediction of Parkinson’s disease.

Fig. 6.

LOSO cross validation results of the proposed MR method for PD voice dataset with imbalanced and balanced data.

Fig. 7.

Per-Fold Accuracy for LOSO Validation.

Table 16 presents the statistical significance (p-values), 95% confidence intervals, and the standard deviation of residuals for select acoustic and nonlinear dynamic features extracted from voice recordings. Most variables demonstrate statistically significant associations (p < 0.05), indicating potential discriminative power between the analyzed groups. Notably, the 95% confidence intervals for variables such as MDVP: Fo(Hz), MDVP: Fhi(Hz), MDVP: Flo(Hz), MDVP: Jitter(%), MDVP: Shimmer, RPDE, DFA, spread2, and D2 do not include zero, reinforcing the reliability of these effects. An exception is HNR, whose confidence interval includes zero, suggesting its contribution may be less robust. The standard deviation of residuals (0.2375) indicates a relatively low level of unexplained variation, suggesting that the model captures the majority of the variability in the data.”

Table 16.

p-value and confidence interval for BOMLR method.

Variable	p-value	95% Confidence interval
MDVP: Fo(Hz)	0.024	[−0.0044, −0.0004]
MDVP: Fhi(Hz)	0.037	[−0.0044, −0.0004]
MDVP: Flo(Hz)	0.041	[−0.0005, −0.0001]
MDVP: Jitter(%)	0.035	[−0.0031, −0.0001]
MDVP: Shimmer	0.029	[−15.612, −0.384]
HNR	0.046	[−0.0381, 0.0165]
RPDE	0.016	[−0.027, −0.001]
DFA	0.044	[−0.825, −0.044]
spread2	0.018	[0.022, 1.135]
D2	0.022	[0.179, 1.925]
Standard Deviation of Residuals:	0.2375

Table 17 summarizes the model’s performance across several key metrics, each reported as the mean value ± standard deviation and the corresponding 95% confidence interval. The high accuracy, F1-score, and ROC-AUC indicate strong classification performance, while the R², RMSE, and MAE values reflect the model’s effectiveness in regression tasks. The narrow confidence intervals suggest consistent and reliable results across repeated evaluations.

Table 17.

Performance metrics of the proposed Model, with Mean ± Standard deviation and confidence Interval.

Metric	Mean ± SD	95% Confidence interval
Accuracy	93.8% ± 1.4%	[91.1%, 96.5%]
F1-Score (macro)	0.921 ± 0.018	[0.887, 0.953]
ROC-AUC (macro)	0.948 ± 0.012	[0.924, 0.968]
R² Score	0.874 ± 0.026	[0.823, 0.925]
RMSE	0.237 ± 0.015	[0.207, 0.266]
MAE	0.188 ± 0.012	[0.164, 0.212]

Table 18 presents the performance of the model across 10 independent training (or validation) runs. For each iteration, the coefficient of determination (R²), root mean squared error (RMSE), and mean absolute error (MAE) are reported. These metrics demonstrate the consistency and variability of the model’s predictive accuracy and error rates, highlighting the robustness of the results across multiple runs. To validate the generalizability and clinical applicability of the proposed methodology, we implemented the same approach on both the PD Voice Dataset (from Kaggle) and the externally sourced PPMI clinical dataset. The description of the PPMI dataset is shown in Table 19.

Table 18.

Performance metrics of the proposed system across 10 Runs.

Iteration (Runs)	R²	RMSE	MAE
1	0.86	0.24	0.19
2	0.88	0.23	0.18
3	0.86	0.21	0.16
4	0.87	0.25	0.20
5	0.86	0.24	0.21
6	0.84	0.23	0.17
7	0.85	0.22	0.19
8	0.87	0.21	0.20
9	0.86	0.24	0.19
10	0.87	0.25	0.18

Table 19.

Description of PPMI dataset.

Category	Details
Total Subjects Enrolled	683
PD (Parkinson’s Disease)	423 untreated PD subjects
HC (Healthy Controls)	196 subjects
SWEDD	64 subjects (Scans Without Evidence of Dopaminergic Deficit)
Study Sites	24 sites
PD Enrollment Criteria	Asymmetric bradykinesia or tremor plus DAT binding deficit on SPECT

As shown in Fig. 8, the model achieved consistently high performance across both datasets, with accuracy values of 0.97 and 0.98, respectively. These results confirm that the proposed method maintains robustness and predictive strength even when applied to real-world clinical data, thereby enhancing the credibility and stability of the model.

Fig. 8.

Performance of the proposed model in PD voice dataset and PPMI dataset.

Table 20 outlines key risks and corresponding strategies when using synthetic data generated through amplitude slicing for Parkinson’s disease prediction. It emphasizes the importance of maintaining the integrity of clinically relevant features while addressing issues like class imbalance, limited variability, and model overfitting. The use of validation techniques and realistic transformations ensures that the model generalizes well without compromising on diagnostic accuracy.

Table 20.

Potential risks and mitigation strategies in using synthetic voice data for parkinson’s disease Prediction.

Aspect	Potential risk	Mitigation strategy
Synthetic Data (Amplitude Slicing)	Overfitting due to learning artifacts or unrealistic patterns	Carefully controlled augmentation preserving pathological features42.
Data Imbalance	Model may become biased toward majority class	Class balancing through augmentation43.
Limited Real Variation	Reduced diversity despite synthetic size increase	Used physiologically plausible transformations only44.
Model Complexity	Risk of memorizing patterns in small datasets (high variance models)	Regularization and cross-validation applied45.

To demonstrate the effectiveness of the proposed CNDWT method in classifying Parkinson’s disease (PD) and healthy subjects using amplitude slicing augmentation, the ROC curve, AUC distribution, and confusion matrix are presented. The ROC curve in Fig. 9 shows excellent classification performance with an AUC of 1.00, indicating perfect separation between the two classes. The confusion matrix in Fig. 10 supports this with 56 out of 59 samples correctly classified. The AUC distribution obtained through Leave-One-Subject-Out (LOSO) validation, shown in Fig. 11, reveals consistently high scores across subjects, with most values close to 1.0 and only a few outliers, confirming the model’s strong generalizability and robustness after augmentation.

Fig. 9.

ROC Curve for PD classification.

Fig. 10.

Confusion Matrix for PD prediction.

Fig. 11.

AUC Distribution using LOSO validation.

Discussion

The proposed CNDWT method for predicting Parkinson’s disease at an early stage uses the cross NDWT method for decomposition and reconstruction of the voice signal, and these signals are amplitude sliced for augmentation. A Bayesian Optimized Multiple Linear Regression Model (BOMLR) is used to predict PD at an early stage. On evaluating the performance of the proposed CNDWT method in conjunction with the BOMLR algorithm, the proposed method improves the prediction and classification of Parkinson’s disease (PD) using voice signal data. A key contribution of this study lies in the effective selection of highly correlated vocal features, which are identified using the correlation matrix shown in Table 6. Features such as MDVP: Jitter (%), MDVP: RAP, and Jitter: DDP exhibited strong correlations, with coefficients above 0.9, indicating a significant influence on PD detection. These relationships are further illustrated in Table 7 through correlation plots, validating the relevance of the selected attributes for building a robust prediction model. The performance of the proposed CNDWT method is compared with several conventional regression and machine learning algorithms, including Random Forest, Support Vector Machines (SVM), Decision Trees, and Multi-Layer Perceptron. As shown in Fig. 5 (A) to (F), the proposed CNDWT method consistently outperforms traditional techniques across various evaluation metrics such as accuracy, precision, Matthews Correlation Coefficient (MCC), ROC curve area, Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R² values. Specifically, the BOMLR model (which integrates CNDWT with multiple regression) achieved higher classification precision and predictive power compared to FFT and DWT-based methods, as evidenced by lower error rates (FNR, FPR, FOR) and improved accuracy and sensitivity. The proposed method consistently outperformed existing regression techniques in terms of accuracy, precision, MAE, RMSE, R², MCC, and ROC values, indicating strong predictive performance and robustness. It also demonstrated superior results compared to traditional feature extraction methods like DWT and FFT, suggesting more effective feature representation. These findings are largely consistent with previous studies that highlight the benefits of advanced or hybrid approaches.

In Table 16, the statistical analysis identified several vocal and nonlinear dynamic features as significant predictors of Parkinson’s disease. Variables such as MDVP: Fo(Hz), MDVP: Fhi(Hz), MDVP: Flo(Hz), Jitter(%), and Shimmer showed statistically significant associations (p < 0.05) with the presence of Parkinson’s, with negative confidence intervals indicating a potential inverse relationship with disease severity. Similarly, nonlinear features like RPDE, DFA, spread2, and D2 also emerged as significant, highlighting the relevance of signal complexity and vocal irregularity in distinguishing affected individuals. The standard deviation of residuals (0.2375) suggests a reasonably good model fit. These findings support previous literature emphasizing the diagnostic value of vocal perturbation and nonlinear dynamics in Parkinson’s prediction. An important aspect addressed in this study is the issue of data imbalance, particularly evident in the original PD voice dataset, which contains only 8 healthy subjects compared to 23 PD-affected individuals. This imbalance initially led to suboptimal model performance, with accuracy and F-score values around 68–70%. However, by applying amplitude slicing augmentation, the class distribution was balanced, resulting in a significant improvement. The final evaluation using Leave-One-Subject-Out (LOSO) cross-validation yielded precision, recall, and F-score values of approximately 98%, confirming the effectiveness of the augmentation approach in addressing skewed data and enhancing classifier performance (Fig. 6).

Table 17 highlights robust model performance, with high mean accuracy (93.8%) and AUC (0.948), and relatively low RMSE and MAE values, confirming strong predictive ability and minimal error. The iteration-wise evaluation of regression performance across 10 runs shown Table 18 indicates stable R² values (ranging from 0.84 to 0.88) and consistent RMSE and MAE scores, validating the reliability of the model. Finally, in Table 20, the risk-mitigation table identifies potential concerns like overfitting, data imbalance, limited variation, and model complexity, and outlines strategies such as physiologically plausible augmentations, class balancing, and regularization to maintain model generalizability and robustness. Figure 20 compares classification performance metrics between the PD Voice Dataset and the PPMI Dataset, showing consistently high accuracy, precision, recall, F1-score, and AUC values for both datasets. Figures 9, 10 and 11 demonstrate that the proposed CNDWT method achieves near-perfect classification between PD-affected and healthy subjects.

The current study employs per-fold validation for internal evaluation. As future work, we plan to perform external validation on newly acquired, independent clinical datasets using rigorous cross-validation techniques. This will provide further evidence of the model’s generalizability across diverse populations. Subsequently, we intend to progress to prospective validation by deploying the model in a live clinical setting, where its predictions will be evaluated on consecutive, unselected patient cases in real time. This phased approach will help establish the robustness and practical applicability of our model before clinical implementation.

Conclusion

Early detection of Parkinson’s disease is necessary to avoid complications for PD-affected patients. Early detection provides an opportunity for patients to receive proper medical treatment and for medical professionals to begin therapy. PD-affected patients mostly exhibit articulation disorders, and therefore the study uses a voice dataset for prediction. In this paper, early detection of PD is achieved using the cross NDWT method, in which decomposition of the input signal is performed using Haar wavelet transform and reconstruction is done using DB3 transform. Initially, the raw input voice data is augmented using amplitude slicing, and new synthetic data are created to balance the dataset. Next, the correlation among the attributes is determined, and the highly correlated attributes are used for PD prediction by applying the Bayesian Optimized Multiple Linear Regression (BOMLR) model. To prove that CNDWT performs better than CDWT, this work compares the performance of both methods, and the results show that CNDWT outperforms CDWT. The performance of the BOMLR method used for classification and prediction of Parkinson’s disease is compared with state-of-the-art methods such as SVM, DT, and random forests, and BOMLR performs better. After evaluating the proposed method against existing methods, the results clearly demonstrate that our proposed CNDWT method outperforms other methods in all cases, achieving an accuracy of 99% for detecting Parkinson’s disease at an earlier stage. In addition to the PD voice dataset, the PSD dataset and the UCINLR dataset are used as input to showcase the efficiency of the proposed CNDWT method in the prediction and classification of PD. The practical implications of this research are significant for early and non-invasive diagnosis of Parkinson’s disease (PD). The proposed method, using voice data, offers a low-cost, accessible, and easily deployable solution for preliminary PD screening, which could be integrated into telemedicine platforms or mobile health applications. This can be particularly beneficial in remote or underserved areas where access to neurologists or advanced diagnostic tools is limited. The high accuracy and reliability of the model suggest it could assist clinicians in monitoring disease progression or flagging cases that require further evaluation. Future research could address the observed anomalies by incorporating multimodal features for Parkinson’s disease prediction rather than relying solely on voice data. Integrating additional data sources, such as EEG signals, gait analysis, handwriting patterns, or sensor-based movement data, could provide a richer and more comprehensive view of the patient’s condition. This multimodal approach may help the model better capture complex and heterogeneous symptoms of PD, potentially improving prediction accuracy, robustness, and reducing variability across different datasets. Exploring fusion strategies for combining these modalities effectively would be a valuable direction for future studies. While the current study uses amplitude slicing to extract features from voice signals, incorporating random slicing could simulate different signal segments and temporal variations that naturally occur in real-world speech. This can enhance the model’s robustness to noise, speaker variability, and recording conditions. Moreover, random slicing may uncover subtle, localized patterns in the voice data that are relevant for early Parkinson’s disease detection. The limitation of the proposed work is the use of a static, cross-sectional voice dataset, which does not capture the temporal progression of Parkinson’s disease. Since PD is a progressive neurodegenerative disorder, relying solely on isolated voice samples may restrict the model’s ability to identify subtle changes that occur over time. This limits the potential for reliable early diagnosis or monitoring of disease progression.

Author contributions

B.Veena, M.S.P.Subathra: conception or design of the work; B.Veena, N.R. Shanker: Data acquisition, analysis, or interpretation of data; B.Veena, S.Thomas George and N.R. Shanker: Creation of software used in the work; N.R Shanker, M.S.P Subathra and S.Thomas George: Drafted the work and Revision of the work.

Funding 

The authors received no specific funding for this study.

Data availability

The dataset is publicly available at: https://www.kaggle.com/datasets/vikasukani/parkinsons-disease-data-set.

Declarations

Competing interests

The authors declare no competing interests.