The dataset was collected from the UCI ML data source which contains 195 speech instances with 24 features [18] . This dataset was generated in 2008 from the University of Oxford, Irvine. The working dataset consists of biomedical speech recordings from persons of various ages, where 147 positive (PD) cases and 48 healthy instances. After that, the “status” features was set into “1” for identifying positive cases and “0” for negative cases. Table 1 summarizes the features of the working dataset. The working dataset was divided into training set (80%) and testing set (20%).

The working dataset was processed through the following steps: outlier identification, data augmentation, and balancing. Identifying outliers is a vital phase in preparing data because unusual data layers could have a negative influence on model performance and dependability [19] . Two methods named Winsor and z-score techniques are applied to identify outliers in data layers. The Z-score measures the standard deviation of a data layer using the average value for easily identifying outliers [20] Winsorization or Winsor method [21] is applied for identifying outliers that involves substituting very high or low values of data with fewer severe ones. Instead of eliminating outliers directly, this approach substitutes high values with the nearest values within a certain percentile range. According to this approach, large numbers might occur as a result of measurement mistakes or random oscillations, rather than being outliers.

Data balancing is the crucial part in ML approach which may distort the findings of the test. To overcome this issue, our experiment used a challenging data balancing strategy known as SMOTE-Tomek. It balances the unbalanced class by integrating SMOTE over-sampling and Tomek under-sampling strategies. It generates simulated specimens in feature set. It chooses a minimal feature set to calculate its k-nearest neighbors [22] .

Data augmentation is a critical operation that enriches datasets by using various methods to increase their quantity and variety [23] . Several augmentation strategies have been utilized in this case. When the Tomek connection is established between two scenarios, then almost every feature from the feature set is deleted. This mechanism is useful when a large number of dataset is needed. Applying revised steps of the working data increases the framework’s resilience. Table 2 shows the data preprocessing stages for processing PD dataset. Figure 2 shows the data distribution chart. The above two techniques (SMOTE-Tomek and data augmentation) are applied after the train/test split to prevent information leakage.

Feature selection approach selects high impactful features from the final data set to boost the efficacy of the suggested method. Several feature selection techniques i.e., mRMR for feature selection, LDA and PCA for feature reduction are used in feature engineering field. Among these techniques, LDA exhibited outstanding performance. The working principle of this technique is described in below.

The dimensionality reduction approach named LDA is widely utilized in feature reduction and selection problems. It aims to identify the linear combination of best features (LCBF) from the entire dataset. It transforms the higher dimensional-feature space (DFS) into a lower DFS based on the LCBF [24] . It improves the ratio of class variance, resulting in the optimum routes in the LCBF for future differentiation [25] . It boosts class reparability, integrates classified data for supervised training, and enhances the effectiveness of classification, especially in cases when classes are well divided [26] .

Let S be the dataset with f features and n instances that is partitioned into m classes. The output of the feature set F is defined by Equation (1).

$F_j=\frac{1}{I_j}\underset{j:u_j=m}{{\displaystyle \sum }}{v}_j$ (1)

Where, $I_j$represents the instances for feature m, $u_j$indicates class label for instance j, and $v_j$ indicates feature map for instance j. The class scatters inside ( $L_{in}$) and outside ( $L_{out}$) the feature map are defined by Equation (2) and Equation (3), respectively.

$L_{in}=\underset{j=1}{\overset{n}{{\displaystyle \sum }}}{I}_j\left(F_j-F\right){\left(F_j-F\right)}^t$ (2)

$L_{out}=\underset{j=1}{\overset{n}{{\displaystyle \sum }}}\underset{j:u_j=m}{{\displaystyle \sum }}\left(u_j-F_m\right){\left(u_j-F_m\right)}^t$ (3)

Three ensemble ML models named boosting, voting, and stacking are applied in this work to predict PD. Each ensemble ML model is the combination of multiple ML classifier which boosts the prediction accuracy by leveraging the power of multiple ML models [27] . By leveraging the variety of the basic classifiers, each model can capture many data attributes while limiting the risk of overfitting [28] . Among these models, stacking provides best results. The working mechanism of this model is described below.

Stacking is an ensemble ML model that combines several base classifiers (BCL) [29] and trained them to produce unique prediction. Then these unique predictions are stacked and transmitted into a meta-classifier (MTC) [30] to produce final prediction. The final prediction of this model is defined by Equation (4). The main concept underlying stacking is to train the base predictors on a single dataset concurrently. In working dataset, the training samples are represented by S and the predicted outcome is $p_k$; where k indicates the number of classifiers. Lastly, the predicted results are represented by Equation (4).

$\hat{p}=\underset{k=1}{\overset{S}{{\displaystyle \sum }}}{p}_k\left(q\right)$ (4)

These predictions are used as input for the MTC. The output of the MTC for data point q is calculated by vector P ̂(q)and defined by Equation (5). The desired result p ̂is obtained by applying the MTC named LR [31] on the BCL output P ̂(q). The final outcomes are calculated by Equation (6).

$\hat{P}\left(q\right)=\left[{\hat{p}}_1\left(q\right),{\hat{p}}_1\left(q\right),\cdot \cdot \cdot ,{\hat{p}}_s\left(q\right)\right]$ (5)

$\hat{p}=f\left(\hat{P}\left(q\right)\right)$ (6)

The current investigation averages the expected likelihoods of each category and chooses the class with the greatest mean possibility as the end result. A unique technique is utilized to evaluate a MTC on the BCL projections in order to reach the ultimate choice, with the goal of learning how to optimally integrate the assumptions. Figure 3 depicts the entire architecture of the suggested framework concept used in this article. Algorithm (1) illustrates how the offered frame work produces final result step by step. The incorporation of numerous models into the suggested model may raise computing demands and make the system more difficult to administer.

To deal with this drawback, four feature analysis techniques are utilized to minimize the complexity of the model by lowering the dimensionality of the feature space and simplifying the framework. LDA determines the optimum linear combinations to distinguish classes, PCA reduces the initial attributes to make an optimal feature set and mRMR chooses the greatest number of significant and minimal redundant characteristics. By lowering the number of characteristics, these strategies result in quicker computing, a decreased risk of excessive fitting, and more comprehensible scenarios, eventually making the algorithm more effective and manageable.

To develop a dependable and transparent system for PD diagnosis, it is critical to graphically display and explain how ML models make judgments [6] . The use of interpretable AI approach is crucial for verifying the final decision of the ML classifier [32] [33] . This research reflects two interpretable methods’ named LIME and SHAP that increase the accessibility and interpretation of the final results.

SHAP is a collaborative theory-based strategy to interpret the final outcomes of the ML classifier [34] . In collaborative strategy, the Shapley result is a way for evenly distributing the “payout” among individuals on the basis of marginal contribution. In this strategy, each feature in the prediction is treated as a player, and the SHAP score measures its individual contribution by comparing the model’s output with and without that feature, relative to the average prediction [35] . SHAP evaluates all conceivable feature subsets (coalitions) and estimates each feature’s contribution to the probability through evaluating the modification in prognosis when the attribute is added to the coalition.

The Shapley value ${\delta }_m$ for m^th feature quantifies the average feature contribution of the model’s prediction and considers all probable subsets of features that is defined by Equation (7). In this context, y(w) denotes the feature map for feature w and it can be calculated as the difference between the final estimation and the baseline estimation of the proposed model. The possible feature set is represented by 2^k for all k. To compute a Shapley value, SHAP evaluates the marginal contribution of feature m across these subsets and then averages the results over all possible permutations. As shown in Equation (8), the model’s outcome for a specific input p is expressed as the sum of a baseline value ${\delta }_0$ and the individual contributions of all attributes. The final outcome from the SHAP algorithm is denoted by $\hat{g}\left(p\right)$ and calculated by Equation (8).

${\delta }_m=\underset{w\subseteq k-\left\{m\right\}}{{\displaystyle \sum }}\frac{\left|w\right|!.\left(\left|k\right|-\left|w\right|-1\right)!}{\left|k\right|!}\left(y\left(w\cup \left\{m\right\}\right)-y\left(w\right)\right)$ (7)

$\hat{g}\left(p\right)={\delta }_0+\underset{m=1}{\overset{k}{{\displaystyle \sum }}}{\delta }_m$ (8)

This mechanism emphasizes the projected outcomes for particular scenarios above providing a thorough understanding of the system across the entire dataset [36] . Applying the LIME methodology offers useful knowledge into how many aspects affect PD, distinguishing those that participate positively from those that have a negative effect. LIME aims to interpret the selected features (represented by Z) by analyzing the local vicinity of that feature. To achieve this, it generates a set perturbed samples z’, which are slight disparity of z. These perturbed samples form a neighborhood around z. Each sample z’ is allocated a proximity weight Ꞷ(Z’) which reflects its similarity to z. From this set, a subset z’ϵ Z of perturbed samples is chosen, and each is weighted based on its closeness to z, as calculated in Equation (9). Then an explainable network $M_L$ is examined on these weighted samples locally to find out the estimated predictions of the local model $M_L$. These local predictions are calculated by Equation (10).

$\omega \left(Z^{\prime }\right)=\frac{\pi \left(z^{\prime }\right)}{{{\displaystyle \sum }}_{{z^{\prime }}_k\in {Z^{\prime }}_N}\pi \left({z^{\prime }}_k\right)}$ (9)

$M_I\left(z\right)=\arg {\min }_z\underset{{z^{\prime }}_k\in {Z^{\prime }}_N}{{\displaystyle \sum }}\omega \left({z^{\prime }}_k\right).L_f(M_L\left(z\right),M_L\left({z^{\prime }}_k\right)+\delta \left(M_I\right)$ (10)

In Equation (10), $L_f$ is the loss function that is measured using the difference between $M_L$ and $M_I$ over the local neighborhood, and $\delta \left(M_L\right)$ is the regularization parameter of the interpretable model $M_I$.

Firstly, we evaluate three ML models without feature selection technique that is shown in Table 4 . Table 4 indicates that the stacking network provided high classification rate. We also show that the accuracy of the voting model is lower than the stacking network. However, we trained these ML models with three feature selection algorithms to increase the classification rate. The experimental results of all ML models with three feature selection techniques are demonstrated in Table 5 . Table 5 reflects that the stacking EML model with LDA feature selector provides best results. The voting ML model with mRMR shows the lowest accuracy than other models. On the other hand, boosting and stacking models are provided same and highest precision of 100% using LDA feature selector. In Table 5 , the boosting model with PCA provides lowest recall value of 85.26% and the voting model with mRMR provides lowest F1 value of 85.72%.

In this experiment, we implement k-fold (k = 5) cross-validation on the PD dataset to assess the model’s robustness and generalizability. The PD dataset is fold into 5 parts and then trained each part using the best Stacking model. Table 6 reflects the experimental results of the Stacking model with 5-fold cross validation.

In this experiment, diverse statistical metrics like classification report, ROC curve, confusion metric, and AUC-ROC curve for all ML models are shown to evaluate the proposed stacking + LDA model. Figure 4 represents different evaluation metrics of the proposed model.

The LDA was fitted on the training portion of the dataset instead of refitting on the full dataset before testing to avoid information leakage issue. This was ensured by placing the model within the cross-validation loop. At no point was the model trained on or exposed to the test data prior to evaluation.

Recently, many authors have developed PD disease prediction system applying four ML models. Though their proposed system was provided outstanding results, but their proposed system has some pitfalls like data imbalances and irregularly optimum characteristics set. That’s why three feature selection methods are employed to mitigate these pitfalls. We compare our proposed work with the previous works in Table 7 .

The proposed stacking model demonstrates competitive performance compared to state-of-the-art deep learning (DL) approaches for PD detection, particularly when evaluated using k-fold cross-validation to ensure robustness and generalizability. Unlike deep models such as CNNs, LSTMs, or Transformers, the stacking model is lightweight, easy to deploy, and computationally efficient. DL models require large datasets, significant computational resources and complex implementation, where our proposed system is suitable for real-time applications. While DL models may achieve slightly higher accuracy on large datasets, the stacking model maintains high predictive accuracy on smaller datasets with lower training time and inference cost. Moreover, it offers better model interpretability through XAI tools, which is crucial for clinical decision support. This makes the proposed approach not only effective but also practical for healthcare deployment scenarios where transparency, speed, and cost-efficiency are essential.

The output of the SHAP algorithm for the PD features are shown in Figure 5 . In Figure 5 , x-axis shows the predicted SHAP scores that impact on framework output and y-axis indicates the overall importance of individual characteristic. Each dot represents feature contributions for specific data instances, with their position along the x-axis showing the extent of the influence. The color of each dot reflects the actual value of the characteristic—red for max values and blue for min values—highlighting how different value ranges influence the estimation. This visualization offers a strong interpretable overview of which features drive the method’s decisions.

Figure 6 reflects the bar chart output of the SHAP algorithm. In Figure 6 , x-axis shows the mean absolute SHAP value and y-axis lists the individual features displaying the mean SHAP values of different features. From Figure 6 , we see that three features named MDVP: Fo (Hz), PPE, and spread1 have the highest SHAP value where the MDVP: jitter (%) has the lowest SHAP value (+0.01). The bar labeled “Sum of 13 other features” reflects the combined contribution of less influential features. This plot provides a clear summary of which features are most significant in driving the model’s decisions.

Figure 7 reflects the waterfall output of the SHAP algorithm. The base value (E[f(x)] = 0.737E [f(x)] = 0.737E [f(x)] = 0.737) represents the average model output across all data. In Figure 7 , the red arrow indicates that the prediction of the features are increased, while blue arrow represents that the prediction of the features are decreased. For example, features like “PPE”, “spread1”, and “MDVP: Fo(Hz)” pushed the prediction higher, while “MDVP: Jitter(%)” slightly decreased it. This plot provides an interpretable breakdown of how specific feature values influenced the final prediction for one particular data point.

Figure 8 reflects a decision plot of a ML model using LIME algorithm. In Figure 8 , the bar chart at the top left shows the final predicted probabilities—0.08 for PD and 0.92 for Non-PD. The center portion contains a tree-based breakdown of decision paths, where each branch shows how specific feature thresholds influence the prediction toward either class. Each decision node includes the feature, its condition, and the corresponding contribution to the framework’s prediction. On the right, a summary table lists the actual results of the best characteristic used in the estimation. Most feature contributions in this example push the prediction toward the Non-PD category, aligning with the high Non-PD probability score. This visualization provides a step-by-step explanation of how individual features and their thresholds contributed to the final classification decision.