Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6, 2591-2608 2025 Publisher: Learning Gate DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate © 2025 by the authors; licensee Learning Gate History: Received: 3 April 2025; Revised: 5 June 2025; Accepted: 10 June 2025; Published: 30 June 2025 * Correspondence: yafaheju@hotmail.com Fuzzy model for the classification of Parkinson’s disease based on voice signals Yamid Fabián Hernández-Julio1,4*, Martha Janeth Prieto-Guevara2, Leonardo Antonio Díaz-Pertuz1, Benjamín Castillo-Osorio1, Mauricio Barrios-Barrios3, Wilson Nieto-Bernal4 1Facultad de Ciencias Económicas, Administrativas y Contables, Universidad del Sinú Elías Bechara Zainúm, Mon-tería, Córdoba 230001, Colombia; yafaheju@hotmail.com (Y.F.H.J.). 2Departamento de Ciencias Acuícolas–Medicina Veterinaria y Zootecnia (CINPIC), Universidad de Córdoba, Monte-ría, Córdoba 230001, Colombia; mprieto@correo.unicordoba.edu.co (M.J.P.G.). 3Computer Science and Electronics Department, Universidad de la Costa, Barranquilla 080001, Colombia; mauri- cio.barrios@gmail.com (M.B.B.) 4Systems Engineering Department, Universidad del Norte, Puerto Colombia, Atlántico 080001, Colombia; wnie- to@uninorte.edu.co (W.N.B.) Abstract: This study presents a Mamdani-type fuzzy logic model for classifying Parkinson’s disease (PD) based on voice signals. The model demonstrates improved performance compared to several existing methods, achieving 97.2% accuracy, 0.9696 sensitivity, 1.0 specificity, and an F-measure of 0.98. These metrics suggest that the proposed model offers higher classification precision than previous approaches. By leveraging fuzzy logic, the model enhances interpretability and addresses some uncertainties inherent in medical data. While the results are promising, further validation with more extensive and diverse datasets is necessary before the model can be integrated into clinical decision support systems for the early diagnosis of PD. Keywords: Parkinson’s disease; fuzzy model. 1. Introduction Parkinson’s disease (PD) is a progressive neurodegenerative disorder that affects millions of individuals worldwide. It is characterized by motor symptoms such as tremors, bradykinesia, rigidity, and non-motor symptoms, including disturbances in the sense of smell, sleep problems, depression, cognitive decline, and voice impairments [1-5]. Recent studies have shown that vocal symptoms, such as dysphonia and changes in pitch, are among the earliest indicators of PD, often preceding noticeable motor impairments [5]. Notably, approximately 90% of PD patients exhibit vocal problems in the early stages, with symptoms ranging from stuttering to deterioration in vocal quality [6, 7]. The early detection of PD is critical for managing its progression and improving patient outcomes. Traditional diagnostic methods rely heavily on clinical assessments, which are often subjective and can miss early signs of the disease. Researchers have proposed a range of non-invasive methods for early detection, including the acoustic analysis of voice signals, physiological signals, and gait analysis [2, 8, 9]. These approaches provide insights into the disease's progression and reduce the need for frequent physical clinical visits, thus easing the clinicians’ workload [6]. Given the significance of vocal symptoms in PD, telemedicine studies have increasingly focused on verbal disorder-based systems [10, 11]. Speech processing, particularly the non-invasive detection of anomalies in physiological speaking, has emerged as a promising method, with features like Jitter, Shimmer, and Fundamental Frequency being pivotal in PD studies [12]. Automated tools that analyze voice signals offer a non-invasive, cost- effective method for early detection. However, many existing machine learning (ML) methods used for 2592 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate voice-based PD classification, such as Support Vector Machines (SVMs) and neural networks, suffer from a lack of interpretability, often referred to as the "black box" problem [7, 13, 14]. This limits their clinical applicability, as healthcare professionals require transparent and understandable models to inform their decisions. In contrast, fuzzy logic provides a promising solution by combining rule-based reasoning with the ability to handle uncertainty and imprecision in medical data. A Mamdani-type fuzzy logic system, in particular, offers a transparent decision-making process through linguistic rules derived from expert knowledge. This study introduces a Mamdani-type fuzzy logic model designed to classify PD using voice signals. The model aims to address the limitations of existing ML approaches by enhancing interpretability while maintaining high classification accuracy. The key contributions of this work are: 1. Development of a fuzzy logic model that outperforms several existing methods regarding accuracy, sensitivity, and specificity. 2. Introduction of an interpretable rule-based system that provides insights into decision-making. 3. A comprehensive evaluation of the model's performance using well-established metrics and comparisons with state-of-the-art methods. In the following sections, we review existing literature on PD classification using voice signals, describe the methodology and design of the fuzzy logic system, and present a detailed evaluation of the model’s performance in comparison to existing benchmarks. 2. Literature Review The classification of Parkinson’s Disease (PD) based on voice signals has gained increasing attention in recent years due to its non-invasive nature and potential for early detection. Several methods have been proposed, ranging from traditional machine learning (ML) algorithms to more complex deep learning approaches. However, many of these methods face limitations that reduce their clinical applicability, particularly in terms of interpretability and handling uncertainty in medical data. 2.1. Machine Learning Approaches for PD Classification The study and classification of Parkinson's disease (PD) have been an area of extensive research. Among the multiple studies conducted, several have employed the UCI dataset for their experiments. Sakar, et al. [6] explored the applicability of the Tuneable Q-factor Wavelet Transform (TQWT) on voice signals of PD patients for feature extraction. Their study juxtaposed the efficacy of TQWT against traditional voice signal processing techniques for PD classification. The findings were promising, with TQWT outperforming other methods by achieving a peak accuracy of 0.86 using the SVM-RBF classifier on voice recordings from 252 participants, validated using the Leave-one-subject- out technique. Akyol [15], on the other hand, ventured into deep learning, employing a Deep Neural Network (DNN) with 753 features. The dataset and specific results for this approach are detailed in the discussion section. Similarly, Xiong and Lu [16] used multiple machine learning methods, including Logistic Regression (LR), Support Vector Machines (%SVM), and Random Forests (RF). The results were validated using 10-fold cross-validation, and the best classification accuracy for all datasets was 0.76. Another noteworthy contribution is from Grover, et al. [7], who focused on forecasting the severity of PD utilizing deep neural networks. Their method was applied to the Parkinson’s Telemonitoring Voice Data Set from UCI, achieving a classification accuracy of 94.4422% for training datasets and 62.7335% for test datasets. Zainudin, et al. [17] adopted a different approach, using radial basis function networks for PD classification. Their study was based on the UCI dataset involving 500 PD patients, and R2 values of 0.7450 and 0.970 were found for multiple linear regression and radial basis function, respectively. 2593 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate Lastly, Hariharan, et al. [2] presented a comprehensive approach, proposing hybrid intelligent systems for PD classification. They integrated techniques like Principal Component Analysis (PCA) and linear discriminant analysis (LDA) for feature pre-processing, followed by LS-SVM, PNN, and GRNN classification. Their combined approach achieved a remarkable classification accuracy of 100% on a dataset with 31 individuals. 2.2. Fuzzy Logic in Medical Applications In contrast, fuzzy logic systems have been used in various medical applications because they can model uncertainty and provide interpretable decisions [18-20]. Fuzzy systems allow for reasoning in the form of linguistic rules, making the decision-making process transparent. Fuzzy logic presents a distinct advantage in clinical decision-making due to its capacity to handle uncertain and vague data, often characteristic of patient symptoms [21-23]. By allowing degrees of truth rather than binary logic, fuzzy logic systems effectively interpret symptoms that are not clear-cut. This flexibility is enhanced by using linguistic "if-then" rules, which provide transparency and interpretability crucial in medical practice. Such interpretability fosters trust among clinicians, who can easily understand and validate the decision-making process [24, 25]. Moreover, fuzzy logic-based systems, especially when integrated into Clinical Decision Support Systems (CDSS), improve decision-making by combining data-driven models with expert knowledge, offering nuanced diagnoses and treatment suggestions that are easy to follow and explain [18, 20, 26]. Recent studies highlight the potential of fuzzy logic in the diagnosis of Parkinson's disease by offering a more nuanced and interpretable classification method compared to traditional machine- learning approaches. Fuzzy logic's ability to handle uncertainty and imprecise data makes it ideal for medical conditions like Parkinson's, where symptoms can be gradual, subjective, and difficult to quantify [27-30]. Despite its strengths, fuzzy logic has been underutilized in PD classification based on voice data. While fuzzy models have shown promise in handling ambiguous medical data, limited research has applied this technique to PD [27-29]. This study aims to fill this gap by developing a Mamdani-type fuzzy logic model that provides high accuracy and makes the decision process interpretable through a system of fuzzy rules. 2.3. Limitations of Existing Methods The primary limitation of many existing machine learning approaches is their inability to offer insight into how predictions are made, which is essential for clinical trust and acceptance. Additionally, models like SVMs and DNNs often require large datasets for training, and their performance can degrade when faced with noise or uncertainty in the data. In contrast, fuzzy logic systems are inherently designed to handle such uncertainty, making them particularly suitable for medical applications where data can be noisy or incomplete. In light of these comprehensive studies, our work aims to offer a fresh perspective, intertwining fuzzy logic to enhance the classification and understanding of Parkinson’s disease through voice signals. 3. Material and Methods The methodology used in this study follows the structured framework proposed by Hernández- Julio, et al. [20]. This framework integrates a step-by-step process for designing and implementing data-driven decision support systems, focusing on iterative development and knowledge-based rules. Below, we outline each of the relevant steps in the context of this research. For this study, more than one thousand fuzzy inference systems were developed. Ultimately, the model with the best performance (Classification accuracy) was selected, and the results were validated with that model. 2594 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate 3.1. Domain Understanding and Gap Identification (Steps 1-3) The first phase of the methodology focuses on gaining a comprehensive understanding of the problem domain, Parkinson’s Disease (PD), and identifying gaps in existing voice-based classification models. An extensive literature review revealed several challenges in current PD classification approaches, such as the lack of model interpretability and the inability to handle uncertainty in the data. These limitations highlight the need for a model that is both interpretable and capable of managing data uncertainty, motivating this study's adoption of fuzzy logic. Figure 1. The proposed five-layer architecture framework. As illustrated in Figure 1, the framework encompasses a five-layer architecture. The entire implementation process of the Fuzzy Model is detailed in Figure 2. Consisting of eleven activity steps, the framework's comprehensive breakdown is as follows: 2595 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate Figure 2. The proposition of the Fuzzy Model implementation. 2596 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate The dataset used in this study was obtained from the UC Irvine Machine Learning Repository [6]. It consists of voice recordings of 188 patients with PD (107 men and 81 women) aged 33 to 87 (65.1±10.9) at the Department of Neurology at Cerrahpaşa Faculty of Medicine, Istanbul University. The control group comprises 64 healthy individuals (23 men and 41 women) aged between 41 and 82 (mean age 61.1 ± 8.9). Each subject phonated the vowel "a" three times, with recordings taken using a microphone set to a 44.1 kHz sampling rate. The dataset features 756 instances with 754 attributes - 753 input variables and one output variable. Importantly, this dataset has no missing values, and all attributes are of integer or real type. This dataset, donated in November 2018, presents a classification challenge. [6]. The dataset contains a rich collection of features, summarized in Table 1. The baseline features of the dataset are visually represented in Figure 3. Table 1. Summary of the feature collections utilized in the research (excluding TQWT). Feature set Measure Explanation # of features Baseline features Jitter variants Variations in jitter are utilized to detect the irregularities present in the vibrating pattern of the vocal cords. This subset of features quantifies the fluctuations in fundamental frequency from one vocal fold cycle to the next. 5 Shimmer variants Shimmer variations are also utilized to capture the vibratory pattern of the vocal cords. In this case, this subset of features quantifies the fluctuations in amplitude from one vocal fold cycle to the next. 6 Fundamental frequency parameters The frequency of vibration of the vocal folds was examined. This frequency's mean, median, standard deviation, minimum, and maximum values were employed for analysis. 5 Harmonicity parameters Because of incomplete closure of the vocal folds, speech pathologies often lead to increased noise components. Parameters like Harmonics to Noise Ratio and Noise to Harmonics Ratio were utilized as features, which help quantify the balance between signal information and noise in the speech. 2 Recurrence Period Density Entropy (RPDE) RPDE gives information about the ability of the vocal folds to sustain stable vocal fold oscillations and quantifies the deviations from F0. 1 Detrended Fluctuation Analysis (DFA) DFA quantifies the stochastic self-similarity of the turbulent noise. 1 Pitch Period Entropy (PPE) PPE measures the impaired control of fundamental frequency F0 using a logarithmic scale. 1 Time- frequency features Intensity Parameters Intensity is connected to the power of the speech signal, typically measured in decibels (dB). This study employed the mean, minimum, and maximum intensity values as features. 3 Formant Frequencies Frequencies enhanced by the vocal tract, known as formants, were utilized as features in this study. The first four formants were explicitly selected for analysis. 4 Bandwidth The first four bandwidths were utilized as features within the frequency range spanning the formant frequencies. 4 Mel Frequency Cepstral Coefficients (MFCCs) MFCCs MFCCs are used to detect the impacts of Parkinson's disease on the vocal tract, distinct from its effects on the vocal folds. 84 2597 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate Wavelet Transform based Features Wavelet transform (WT) features related to F0 WT features quantify the deviations in F0 182 Vocal fold features Glottis Quotient (GQ) GQ provides insights into the durations of glottis opening and closing, indicating the regularity in glottis movement. 3 Glottal to Noise Excitation (GNE) GNE measures the level of turbulent noise resulting from inadequate closure of the vocal folds within the speech signal. 6 Vocal Fold Excitation Ratio (VFER) VFER calculates the quantity of noise generated due to abnormal vocal fold vibration, utilizing nonlinear energy and entropy principles. 7 Empirical Mode Decomposition (EMD) EMD decomposes a speech signal into an elementary signal 6 Figure 3. Image of the baseline features dataset. 3.2. Initial System Design (Steps 4-5) Following domain analysis, the next phase involved designing the initial system architecture. This phase includes defining the features to be extracted from the voice data, as well as the preprocessing steps: Feature Selection: Acoustic features such as jitter, shimmer, harmonic-to-noise ratio (HNR), Baseline, and Mel Frequency Cepstral Coefficients (MFCCs) were selected based on their relevance in identifying voice impairments linked to PD. 3.2.1 Data Preprocessing: Data Cleaning: Noisy and irrelevant data were filtered to improve model performance. 2598 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate Normalization: Feature values were normalized to a standard scale to prevent biases during model training. Data Splitting: There were two types of data-splitting methods. Random sampling: The dataset was split into training and testing sets in different ratios (%): 50-50, 70-30, 80-20, 100-0, ensuring a robust evaluation of the model’s performance. The other method used was the Cross-Validation with k = 10 folds. 3.3. Iterative Design and Development (Steps 6-9) This phase focuses on the iterative design and construction of the decision support system, primarily focusing on the fuzzy logic model. 3.3.1. Knowledge Database Creation (Steps 7–8): A knowledge database was built using voice features systematically analyzed for their relevance in distinguishing between PD and healthy individuals. Pivot Tables: Pivot tables were used to analyze the relationships between the acoustic features and the binary classification outcome (PD vs. healthy). This analysis was the foundation for the feature selection process, identifying which features contributed the most to the classification. 3.3.2. Rule Base Creation (Step 9): A knowledge rule base was constructed based on the insights gained from the pivot table analysis. Fuzzy if-then rules were generated to model the relationships between voice features and the likelihood of PD. The knowledge rule base forms the core of the fuzzy inference system, allowing the model to make transparent decisions based on these rules. 3.4. Implementation and Evaluation of the Fuzzy Logic Models (Steps 10-11) The next phase involves the implementation and evaluation of the Mamdani-type fuzzy logic model: 3.4.1. Mamdani-Type Fuzzy System The fuzzification process converted numerical values of the voice features into linguistic variables such as “Very low,” “Low,” “Moderately low,” “Medium-low,” “Medium-high,” “Moderately high,” “High,” and “Very high.” Triangular and trapezoidal membership functions defined the boundaries of these linguistic variables. The inference engine applied the fuzzy if-then rules to combine these linguistic variables and generate a fuzzy output representing the likelihood of PD. Finally, the defuzzification process converted the fuzzy output into a crisp classification (either PD or healthy). 3.4.2. Evaluation In this stage, the performance and reliability of the developed fuzzy model are rigorously assessed employing several evaluation metrics, which include: Classification Accuracy Sensitivity Specificity Function Measure Area Under the Curve (AUC) Kappa Statistics A detailed exploration and explanation of these metrics, including their respective mathematical formulations, are delineated in [31]. (Please refer to Appendix A for a thorough exposition of the experiments conducted.) 2599 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate 3.4.2.1. Calculating Classification Accuracy Classification accuracy is a pivotal metric that quantitatively measures the model's ability to classify instances correctly. It is computed as follows: During this phase, the efficacy of the fuzzy model was assessed using the subsequent metrics: classification accuracy, sensitivity, specificity, Function Measure, Area under the curve (AUC), and Kappa statistics. These evaluation metrics are comprehensively described in [31] with their respective formulae (All the experiments are available in Appendix A). The classification accuracy was determined by computing the ratio of the sum of true positives (TP) and true negatives (TN) achieved through the classification algorithms to the total count of occurrences, as defined by the equation (1). 𝐴𝐶𝐶 = 𝑇𝑃 + 𝑇𝑁 𝑇𝑃 + 𝐹𝑃 + 𝐹𝑁 + 𝑇𝑁 Where: TP represents True Positives TN symbolizes True Negatives FP denotes False Positives FN indicates False Negatives The accuracy value (ACC) represents the ratio of the correctly classified instances (positive and negative) to the total cases in the dataset, thereby providing a concise yet comprehensive measure of the model's classification precision. Consequently, this metric provides an overall snapshot of how effectively the fuzzy model echoes the observed classification outcomes, offering a transparent insight into its general predictive capability. 4. Results We validated the best fuzzy model, juxtaposing it with alternative models that similarly aimed to predict the output variable through the interaction of inputs. The performance of the proposed Mamdani-type fuzzy logic model was evaluated using two different methods: 10-fold cross-validation and random sampling (different ratios). The accuracy, measured using R² coefficients, was computed for various feature sets, including VFF, Wavelet Transform (WT), TQWT, MFCCs, Intensities, Baseline, and the combination of all features. The results highlight the robustness of the model across different datasets and feature sets. 4.1. Performance Evaluation Using Cross-Validation (10-Folds) The cross-validation results in Table 2 present the R² accuracy coefficients across different datasets and feature sets. As seen, the model achieved consistently high performance across most datasets, with a maximum R² value of 0.975 for several feature sets, including VFF, WT, TQWT, and MFCCs. The performance remained stable for the training data, with an R² value of 1.0 across most feature sets. In contrast, the test set performance showed some variability, particularly with intensities, with a lower test R² value of 0.638. Table 2. Mean of Accuracy percentage (R² coefficients) for datasets and number of features (Cross-validation 10-folds). Datasets/Features VFF WT TQWT MFCC Intensities BWF Baseline All Features All dataset 0.975 0.975 0.975 0.975 0.640 0.975 0.971 0.975 Training 1.000 1.000 1.000 1.000 0.634 0.999 0.994 1.000 Test 0.744 0.744 0.747 0.744 0.638 0.750 0.762 0.744 (1) 2600 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate These results demonstrate the model's ability to generalize across different feature sets, with a slight drop in performance when using intensities as the primary feature. 4.2. Performance Evaluation Using Random Sampling (80-20%) Table 3 shows the accuracy results obtained using random sampling with an 80-20% train-test split. The model achieved strong performance on the training sets, with R² values 1.0 for most feature sets. However, the test set performance varied slightly, with the highest R² value of 0.861 achieved for VFF and Baseline features. The intensities dataset again showed the lowest performance on the test set, with an R² value of 0.748. Table 3. Accuracy percentage (R² coefficients) for datasets and number of features (Random sampling 80-20%). Datasets/Features VFF WT TQWT MFCC Intensities BWF Baseline All Features All dataset 0.972 0.968 0.971 0.971 0.787 0.968 0.972 0.968 Training 1.000 1.000 1.000 1.000 0.797 1.000 1.000 1.000 Test 0.861 0.841 0.854 0.854 0.748 0.841 0.861 0.841 4.3. Analysis and Comparisons The results indicate that the proposed fuzzy logic model performs consistently well across various feature sets and evaluation methods. Notably: • The VFF, WT, TQWT, and MFCC feature sets showed the highest performance in cross- validation and random sampling methods, with test R² values ranging between 0.744 and 0.861. • The intensities feature set consistently underperformed compared to the other sets, with R² values of 0.640 in cross-validation and 0.748 in random sampling. This suggests that intensities alone may not provide sufficient discriminatory power for PD classification. • The baseline and all-features set performed well, with the test set R² values close to those of VFF, WT, and TQWT, indicating that these feature combinations can offer robust classification performance. The cross-validation results show stable model performance across different training and test splits. In contrast, the random sampling results confirm that the model can generalize well to unseen data, mainly using the Baseline, VFF, TQWT, and MFCC feature sets. Table 4 reveals the number of clusters utilized for each subset or group of features, having derived the values from training with 100% of the dataset. The chief aim was to discern the number of clusters necessary to achieve 100% classification accuracy with a single iteration. Once this value was identified, training could proceed with other percentages. As previously mentioned, the minimum value for training with the complete dataset was two. If the cluster number is five, then prior numbers did not accomplish 100% classification accuracy. Similarly, if the value exceeds ten (as in Intensity-Based), it could not attain 100% classification accuracy. Nonetheless, training occurred with the optimal number calculated using pivot tables (Section 5). Table 4 displays the number of clusters for each input variable for optimal performance in individual datasets. 2601 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate Table 4. Clusters numbers for every input variable for the best performance in individual datasets. Feature Group Number of Clusters WT—182 5 Baseline—21 8 VFF—22 4 TQWT—432 2 All features—753 2 MFFCs—84 2 Bandwidth + Formant—8 10 Intensity-Based—3 27 Note: WT: Wavelet Transformation. VFF: Vocal Fold Features. TQWT: Tuneable Q-Factor wavelet transform. MFFCs: Mel Frequency Cepstral Coefficients. Table 5 and Figure 4 present the performance outcomes of the models for the PD dataset and its corresponding sub-datasets. Table 4 showcases experiment results obtained with individual feature subsets, delineating performance metrics (specificity, sensitivity, precision, recall, F-measure, and Area Under Curve) for all datasets. Figure 4 illustrates the confusion matrices for training (80%) and validation (20%) datasets for the baseline feature group, consisting of 21 features. Class 0 represents healthy individuals, while Class 1 denotes Parkinson’s Disease patients. Table 5. Experiment results were obtained with individual feature subsets. Feature Groups—Feature Numbers Sensitivity Specificity Precision Recall F-Measure Area under Curve Training: 50–50% Testing WT—182 0.8554 1.0000 1.0000 0.8554 0.9392 0.8099 Baseline—21 0.8938 1.0000 1.0000 0.8938 0.9439 0.8255 VFF—22 0.8854 1.0000 1.0000 0.8854 0.9392 0.8099 TQWT—432 0.8854 1.0000 1.0000 0.8854 0.9392 0.8099 All features—753 0.8854 1.0000 1.0000 0.8854 0.9392 0.8099 MFFCs—84 0.8854 1.0000 1.0000 0.8854 0.9392 0.8099 Bandwidth + Formant—8 0.8812 1.0000 1.0000 0.8812 0.9369 0.8021 Intensity-Based—3 0.8489 0.5400 0.8369 0.8489 0.8429 0.6997 Training 70–30% Testing WT—182 0.9400 1.0000 1.0000 0.9400 0.9691 0.9063 Baseline—21 0.9276 1.0000 1.0000 0.9276 0.9625 0.8854 VFF—22 0.9353 1.0000 1.0000 0.9353 0.9666 0.8984 TQWT—432 0.9322 1.0000 1.0000 0.9322 0.9649 0.8932 All features—753 0.9322 1.0000 1.0000 0.9322 0.9649 0.8932 MFFCs—84 0.9353 1.0000 1.0000 0.9353 0.9666 0.8984 Bandwidth + Formant—8 0.9338 1.0000 1.0000 0.9338 0.9658 0.8958 Intensity-Based—3 0.8270 0.5839 0.8901 0.8270 0.8574 0.6716 Training 80–20% Testing WT—182 0.9641 1.0000 1.0000 0.9641 0.9817 0.9453 Baseline—21 0.9696 1.0000 1.0000 0.9696 0.9800 0.9401 VFF—22 0.9641 1.0000 1.0000 0.9641 0.9817 0.9453 TQWT—432 0.9625 1.0000 1.0000 0.9625 0.9809 0.9427 All features—753 0.9608 1.0000 1.0000 0.9608 0.9800 0.9401 MFFCs—84 0.9353 1.0000 1.0000 0.9353 0.9666 0.8984 Bandwidth + Formant—8 0.9592 1.0000 1.0000 0.9592 0.9792 0.9375 Intensity-Based—3 0.8480 0.5876 0.8706 0.8480 0.8591 0.7061 Training 100–0% Testing WT—182 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 Baseline—21 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 VFF—22 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 TQWT—432 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 2602 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate All features—753 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 MFFCs—84 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 Bandwidth + Formant—8 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 Intensity-Based—3 0.8270 0.5839 0.8901 0.8270 0.8574 0.6716 Note: WT: Wavelet Transformation. VFF: Vocal Fold Features. TQWT: Tuneable Q-Factor wavelet transform. MFFCs: Mel Frequency Cepstral Coefficients. Bold values indicate the best performance. Table 6. Matrix confusion of the datasets (training (a) and test (b)) for the baseline feature group, with 21 features. Class 0 means healthy people, and Class 1 means Parkinson’s disease patients. (a) Class 0 Class 1 Class 0 165 0 Class 1 0 440 (b) Class 0 Class 1 Class 0 4 23 Class 1 0 124 In Appendix A, all the fuzzy inference models developed in the experiments are listed. The model with the optimal configuration of the best fuzzy inference system used to predict Parkinson’s disease, including inputs, output, and samples from the knowledge rule base is located in the folder path folder/80-20/21 baseline trained Acc_0.972 1 0.861_#Vars_21__NaVars_more than 20 variables_Parkinson_Outputs.fis. 5. Discussion The performance of the proposed Mamdani-type fuzzy logic model for classifying Parkinson’s Disease (PD) based on voice signals was evaluated using two methods: 10-fold cross-validation and random sampling (80-20%). The results demonstrate the model's effectiveness in achieving high accuracy across multiple feature sets but also reveal specific insights regarding the contribution of different acoustic features. 5.1. Key Findings The fuzzy logic model consistently achieved high R² accuracy coefficients for several feature sets, particularly VFF, WT, TQWT, and MFCC. In both cross-validation and random sampling, these feature sets resulted in test set R² values ranging from 0.744 to 0.861, showing the robustness of these features in distinguishing between PD patients and healthy controls. The VFF and Baseline features sets showed the highest performance in both methods, with a test R² of 0.861 in random sampling and 0.744 - 0.762 in cross-validation, respectively. This suggests that vocal fold frequency characteristics are highly predictive of PD, likely due to their sensitivity to early vocal changes caused by the disease. Similarly, WT and TQWT feature sets also performed strongly, showing that these time-frequency transformation techniques can capture the subtle variations in vocal signals associated with PD. In contrast, the intensities feature set performed poorly compared to the others, with test R² values of 0.638 in cross-validation and 0.748 in random sampling. This indicates that intensity-based features alone are insufficient to classify PD accurately, likely because changes in vocal intensity may be less consistent or pronounced in the early stages of the disease. 5.2. Comparative Analysis of Evaluation Methods When comparing the two evaluation methods (cross-validation and random sampling), the random sampling method showed slightly better performance on the test set, particularly for the VFF and baseline feature sets. The highest test R² value obtained was 0.861 for VFF and baseline under random 2603 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate sampling. At the same time, the cross-validation method resulted in slightly lower values, with a maximum test R² of 0.762 for the baseline feature set. This difference in performance could be attributed to the inherent variability in cross-validation, where the model is tested on different subsets of the data. In contrast, random sampling provides a fixed training and test set, which may lead to more stable results in some instances. However, both methods confirm the robustness of the fuzzy logic model across different feature sets, mainly when using frequency-domain features like VFF and TQWT. 5.3. Implications for Clinical Applications The results of this study have several implications for the potential clinical application of the fuzzy logic model in early PD detection: • Vocal fold frequency characteristics (VFF) and time-frequency transformations (WT, TQWT) effectively detect the subtle vocal impairments associated with PD, making them valuable features for developing non-invasive diagnostic tools. • The high accuracy and stability of the model across multiple evaluation methods suggest that the fuzzy logic system could be implemented in clinical decision support systems to aid in the early detection of PD. The interpretability of the fuzzy logic model also makes it a suitable choice for clinical environments where transparency in decision-making is critical. • While the intensity-based features did not perform as well, their inclusion in combination with more predictive features (e.g., VFF, TQWT) may still provide complementary information, particularly in later stages of the disease when vocal intensity might change more dramatically. To validate our results, we juxtaposed them with various studies from existing literature, with a selection criterion based on their recentness and demonstrated effectiveness in classification tasks using similar Parkinson’s Disease datasets. Reference [32] devised a method intertwining minimum average maximum (MAMa) tree and singular value decomposition (SVD) for future extraction, comprising pre-processing, feature extraction, feature selection, and classification stages. Two cases were employed in the application, each utilizing a different combination of these stages, and 1-NN and k-NN classifiers were identified as providing optimal classification accuracies for Case 1 and Case 2, respectively. Remarkably, our results exhibited higher classification accuracy with fewer features than this study. In another study, Reference [15] scrutinized the influence of neuron numbers and activation functions in a Deep Neural Network (DNN) model, optimizing its performance via a growing and pruning methodology. Although a superior performance on the test dataset was achieved, our algorithms consistently achieved a 100% classification accuracy rate on the training dataset. Furthermore, our results for 80–20% of training datasets surpassed those acquired by the author, even with a reduced feature set. In a different approach, Reference [16] employed six supervised machine learning algorithms, obtaining a maximum classification accuracy of 76% using linear discriminant analysis (LDA) across all datasets. Our results showcased competitive with or superior performance compared to these findings. Moreover, Reference [33] explored the future extraction process using Wrappers feature subset selection and four classification techniques. Despite achieving notable outcomes, our performance metrics results prevailed in comparison. Reference [34] applied a gender-based dataset division using a Simple Logistic hybrid system. Although reasonably accurate results for both genders were obtained, our results demonstrated superior accuracy and AUC values. While Reference [6] adopted the tuneable Q-factor wavelet transform (TQWT) and several machine learning classifiers for PD classification, our proposed method yielded higher metrics without relying on this classifier. 2604 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate Following the processes and stages for designing fuzzy systems as laid out by Hernández-Julio, et al. [20] and Hernández-Julio, et al. [19], this approach was characterized by several distinctions and similarities in comparison to previously cited studies and established frameworks: • Identification of Variables: Both our approach and others identified input and output variables for Parkinson's disease classification using similar datasets. • Formulating Membership Functions: Unlike other methodologies that use various algorithms to define membership functions, the framework proposed by Hernández-Julio, et al. [20] allows for a manual choice of the number of functions via clustering techniques, thereby avoiding reliance on randomness or evolutionary algorithms. • Rule Base Generation: Using pivot tables does not require calculations, random factors, or manual parameters to create the fuzzy rule base, which distinguishes our method from those that involve random weights and objective functions. The simplicity and directness of this technique stand out by focusing on minimizing redundant information. The parameters applied within the framework, such as the choice of input and output variables, selection of clustering algorithms, and data partition method, are straightforward and do not necessitate adjustments for random values, weights, or other variables (Figure 5). Additionally, internal adjustments were confined to those used for clustering methods. The minimalistic approach toward computational demand and our algorithms' accurate and efficient processing contribute to a straightforward comprehension of the rules. 2605 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate Figure 5. Required parameters of the proposed framework. The efficacy of our framework has been demonstrated across varied domains, including Medicine [18, 20]. Bioengineering [34]. Aquaculture [35] and Colombian Business Finances [36] indicating its versatile applicability to diverse problems contingent on data availability. According to the results, the FIS with the best performance revealed optimal performance using the baseline dataset with fewer features (applying random sampling and cross-validation) and allocating 80% for training and the remaining 20% for validation. With a classification accuracy of 97.2%, sensitivity and recall of 0.9696, specificity and precision of 1.0, F-Measure of 0.98, and an AUC of 0.9401 for complete datasets, these results can potentially be harnessed in telediagnosis and telemonitoring systems for preliminary disease detection, thereby mitigating the necessity for frequent clinic visits and alleviating clinician workloads, irrespective of pandemic circumstances. The Fuzzy Model, applicable individually or to patient subsets (Appendix B), is thus presented as a viable tool for Parkinson’s disease classification. 2606 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate 5.4. Limitations and Future Directions Despite the promising results, several limitations of this study should be addressed in future research: • The dataset used in this study is relatively small and consists of voice recordings from a limited population. To ensure the model's generalizability, future work should involve more extensive and diverse datasets, including data from different age groups, genders, and stages of PD progression. • Although the fuzzy logic model showed high accuracy, adding more advanced feature selection techniques could further optimize the model’s performance. For instance, combining pivot table analysis with machine learning-based feature selection could help refine the most important features for classification. • The current study focuses on voice signals as the primary data source. Future research could explore combining voice data with other non-invasive biomarkers, such as gait analysis or handwriting patterns, to develop a more comprehensive early detection system for PD. 6. Conclusions In conclusion, the selected Mamdani-type fuzzy logic model demonstrated strong classification performance for PD detection using voice signals, with VFF, Baseline, and TQWT feature sets showing the highest accuracy. The results underscore the potential of non-invasive voice-based diagnostic tools for early PD detection. Future work should focus on validating the model with larger datasets and exploring the integration of multiple data sources to improve diagnostic accuracy further. Institutional Review Board Statement: The study was conducted according to the principles outlined in the Declaration of Helsinki and was approved by the Clinical Research Ethics Committee of Bahcesehir University. Acknowledgments: The primary author sincerely thanks the Administrative Department of Science, Technology, and Innovation (MINCIENCIAS) of Colombia and Universidad del Norte for providing the doctoral scholarship. Transparency: The authors confirm that the manuscript is an honest, accurate, and transparent account of the study; that no vital features of the study have been omitted; and that any discrepancies from the study as planned have been explained. This study followed all ethical practices during writing. Copyright: © 2025 by the authors. This open-access article is distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/). References [1] Parkinson' s News Today, "Parkinson’s disease statistics," in "Parkinson's News Today," BioNews Services, LLC, Pensacola, FL, vol. 2024 2024. https://parkinsonsnewstoday.com/parkinsons-disease-statistics/ [2] M. Hariharan, K. Polat, and R. Sindhu, "A new hybrid intelligent system for accurate detection of Parkinson's disease," Computer Methods and Programs in Biomedicine, vol. 113, no. 3, pp. 904-913, 2014/03/01/ 2014. https://doi.org/10.1016/j.cmpb.2014.01.004 [3] A. Benba, A. Jilbab, and A. Hammouch, "Using human factor cepstral coefficient on multiple types of voice recordings for detecting patients with parkinson's disease," IRBM, vol. 38, no. 6, pp. 346-351, 2017/11/01/ 2017. https://doi.org/10.1016/j.irbm.2017.10.002 https://creativecommons.org/licenses/by/4.0/ https://parkinsonsnewstoday.com/parkinsons-disease-statistics/ https://doi.org/10.1016/j.cmpb.2014.01.004 https://doi.org/10.1016/j.irbm.2017.10.002 2607 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate [4] D. Montaña, Y. Campos-Roca, and C. J. Pérez, "A diadochokinesis-based expert system considering articulatory features of plosive consonants for early detection of Parkinson’s disease," Computer Methods and Programs in Biomedicine, vol. 154, pp. 89-97, 2018/02/01/ 2018. https://doi.org/10.1016/j.cmpb.2017.11.010 [5] Parkinson's Foundation. "Statistics." Parkinson's Foundation. https://parkinson.org/Understanding- Parkinsons/Statistics (accessed march 21st, 2024). [6] C. O. Sakar et al., "A comparative analysis of speech signal processing algorithms for Parkinson’s disease classification and the use of the tunable Q-factor wavelet transform," Applied Soft Computing, vol. 74, pp. 255-263, 2019/01/01/ 2019. https://doi.org/10.1016/j.asoc.2018.10.022 [7] S. Grover, S. Bhartia, Akshama, A. Yadav, and S. K.R, "Predicting severity of parkinson’s disease using deep learning," Procedia Computer Science, vol. 132, pp. 1788-1794, 2018/01/01/ 2018. https://doi.org/10.1016/j.procs.2018.05.154 [8] K. Nguyen, J. G. M. Rui, B. P. Nguyen, M. C. H. Chua, and Y. O. Yang, "Classification of parkinson’s disease- associated gait patterns," in Research in Intelligent and Computing in Engineering, Singapore, R. Kumar, N. H. Quang, V. Kumar Solanki, M. Cardona, and P. K. Pattnaik, Eds., 2021// 2021: Springer Singapore, pp. 595-606. [9] K. Nguyen et al., "Classification of gait patterns using overlapping time displacement of batchwise video subclips," Singapore, 2021: Springer Singapore, in Research in Intelligent and Computing in Engineering, pp. 99-111. [10] B. Erdogdu Sakar, G. Serbes, and C. O. Sakar, "Analyzing the effectiveness of vocal features in early telediagnosis of Parkinson's disease," PLOS ONE, vol. 12, no. 8, p. e0182428, 2017. https://doi.org/10.1371/journal.pone.0182428 [11] H. J. N. C. Gürüler and Applications, "A novel diagnosis system for Parkinson’s disease using complex-valued artificial neural network with k-means clustering feature weighting method," vol. 28, no. 7, pp. 1657-1666, 2017. [12] Y. Guo and M. Hazas, "Localising speech, footsteps and other sounds using resource-constrained devices," in Proceedings of the 10th ACM/IEEE International Conference on Information Processing in Sensor Networks, 2011: IEEE, pp. 330-341. [13] A. Gegov, U. Kaymak, and J. M. C. Souza. "Special Issue on Deep Fuzzy Models." Computational intelligence Society - IEEE. https://cis.ieee.org/images/files/Publications/TFS/special-issues/Special-Issue-on-Deep-Fuzzy-Models.pdf (accessed March 21st, 2019). [14] D. L. Poole, A. K. Mackworth, and R. Goebel, Computational intelligence: A logical approach. New York: Oxford University Press, 1998. [15] K. Akyol, "Growing and pruning based deep neural networks modeling for effective parkinson's disease diagnosis," Computer Modeling in Engineering & Sciences, vol. 122, no. 2, pp. 619-632, 2020. https://doi.org/10.32604/cmes.2020.07632 [16] Y. Xiong and Y. Lu, "Deep Feature Extraction From the Vocal Vectors Using Sparse Autoencoders for Parkinson’s Classification," IEEE Access, vol. 8, pp. 27821-27830, 2020. 10.1109/ACCESS.2020.2968177 [17] N. F. Zainudin, N. Mohamed, N. A. Aleng, and S. H. A. Rusmili, "Application of radial basis function network on Parkinson data," Jurnal Teknologi, vol. 77, no. 33, 2015. [18] Y. F. Hernández-Julio, H. Muñoz-Hernández, J. D. Canabal Guzmán, W. Nieto-Bernal, R. R. González Díaz, and P. Ponciano Ferraz, "Fuzzy knowledge discovery and decision-making through clustering and dynamic tables: Application in medicine," in Information Technology and Systems. ICITS 2019. Advances in Intelligent Systems and Computing, vol. 918, Á. Rocha, C. Ferrás, and M. Paredes Eds. Quito, Ecuador: Springer, Cham, 2019, pp. 122-130. https://doi.org/10.1007/978-3-030-11890-7_13 [19] Y. F. Hernández-Julio, W. Nieto-Bernal, and H. Muñoz-Hernández, Framework for the development of data-driven mamdani-type fuzzy decision support systems based on fuzzy set theory using clusters and pivot tables, 1 ed. Montería, Colombia: Universidad del Sinú Elías Bechara Zainúm, 2021. [20] Y. F. Hernández-Julio, M. J. Prieto-Guevara, W. Nieto-Bernal, I. Meriño-Fuentes, and A. Guerrero-Avendaño, "Framework for the development of data-driven Mamdani-type fuzzy clinical decision support systems," Diagnostics, vol. 9, no. 2, p. 52, 2019. [21] Y. Zheng, Z. Xu, T. Wu, and Z. Yi, "A systematic survey of fuzzy deep learning for uncertain medical data," Artificial Intelligence Review, vol. 57, no. 9, p. 230, 2024/08/05 2024. https://doi.org/10.1007/s10462-024-10871-7 [22] M. Tanveer et al., "Fuzzy Deep Learning for the Diagnosis of Alzheimer's Disease: Approaches and Challenges," IEEE Transactions on Fuzzy Systems, pp. 1-20, 2024. 10.1109/TFUZZ.2024.3409412 [23] O. P. Singh and M. E. Patil, "Analysis of Ambiguity, Vagueness, Fuzziness, Uncertainty, Possibility and Probability in the Natural Language Semantics with Fuzzy Logic," International Research Journal on Advanced Engineering Hub (IRJAEH), vol. 2, no. 05, pp. 1478-1483, 05/24 2024. 10.47392/IRJAEH.2024.0204 [24] L. Magdalena, "Fuzzy Systems Interpretability: What, Why and How," in Fuzzy approaches for soft computing and approximate reasoning: Theories and applications: Dedicated to bernadette bouchon-meunier, M.-J. Lesot and C. Marsala Eds. Cham: Springer International Publishing, 2021, pp. 111-122. https://doi.org/10.1007/978-3-030- 54341-9_10 https://doi.org/10.1016/j.cmpb.2017.11.010 https://parkinson.org/Understanding-Parkinsons/Statistics https://parkinson.org/Understanding-Parkinsons/Statistics https://doi.org/10.1016/j.asoc.2018.10.022 https://doi.org/10.1016/j.procs.2018.05.154 https://doi.org/10.1371/journal.pone.0182428 https://cis.ieee.org/images/files/Publications/TFS/special-issues/Special-Issue-on-Deep-Fuzzy-Models.pdf https://doi.org/10.32604/cmes.2020.07632 https://doi.org/10.1007/978-3-030-11890-7_13 https://doi.org/10.1007/s10462-024-10871-7 https://doi.org/10.1007/978-3-030-54341-9_10 https://doi.org/10.1007/978-3-030-54341-9_10 2608 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 6: 2591-2608, 2025 DOI: 10.55214/25768484.v9i6.8453 © 2025 by the authors; licensee Learning Gate [25] J. M. Alonso, C. Castiello, and C. Mencar, "Interpretability of fuzzy systems: Current research trends and prospects," in Springer Handbook of Computational Intelligence, J. Kacprzyk and W. Pedrycz Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2015, pp. 219-237. https://doi.org/10.1007/978-3-662-43505-2_14 [26] S. Paul, K. S. Ray, and D. Saha, "Clinical decision support system using fuzzy logic programming and data analysis," in Emerging Technologies in Data Mining and Information Security. Lecture Notes in Networks and Systems, Singapore, J. M. R. S. Tavares, S. Chakrabarti, A. Bhattacharya, and S. Ghatak, Eds., 2021, vol. 164: Springer Singapore, in Emerging Technologies in Data Mining and Information Security, pp. 175-183. [27] H. Azadi, M. R. Akbarzadeh-T, H. R. Kobravi, and A. Shoeibi, "Robust Voice Feature Selection Using Interval Type- 2 Fuzzy AHP for Automated Diagnosis of Parkinson's Disease," IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 2792-2802, 2021. 10.1109/TASLP.2021.3097215 [28] M. Nilashi et al., "Accuracy analysis of type-2 fuzzy system in predicting parkinson’s disease using biomedical voice measures," International Journal of Fuzzy Systems, vol. 26, no. 4, pp. 1261-1284, 2024/06/01 2024. https://doi.org/10.1007/s40815-023-01665-0 [29] H. Ahmadi, L. Huo, G. Arji, A. Sheikhtaheri, S.-M. J. B. Zhou, and B. Engineering, "Early diagnosis of Parkinson’s disease using a hybrid method of least squares support vector regression and fuzzy clustering," vol. 44, no. 3, pp. 569- 585, 2024. [30] R. Khatwad, S. Tiwari, Y. Tripathi, A. Nehra, and A. Sharma, "Parkinson's disease detection using voice and speech— systematic literature review," in Bio‐Inspired Optimization for Medical Data Mining, S. Srivastava, A. Anand, A. Kumar, B. Saini, and P. S. Rathore Eds., 2024, pp. 41-74. https://doi.org/10.1002/9781394214211.ch3 [31] A. Onan, "A fuzzy-rough nearest neighbor classifier combined with consistency-based subset evaluation and instance selection for automated diagnosis of breast cancer," Expert Systems with Applications, vol. 42, no. 20, pp. 6844-6852, 2015/11/15/ 2015. https://doi.org/10.1016/j.eswa.2015.05.006 [32] T. Tuncer, S. Dogan, and U. R. Acharya, "Automated detection of Parkinson's disease using minimum average maximum tree and singular value decomposition method with vowels," Biocybernetics and Biomedical Engineering, vol. 40, no. 1, pp. 211-220, 2020/01/01/ 2020. https://doi.org/10.1016/j.bbe.2019.05.006 [33] G. Solana-Lavalle, J.-C. Galán-Hernández, and R. Rosas-Romero, "Automatic Parkinson disease detection at early stages as a pre-diagnosis tool by using classifiers and a small set of vocal features," Biocybernetics and Biomedical Engineering, vol. 40, no. 1, pp. 505-516, 2020/01/01/ 2020. https://doi.org/10.1016/j.bbe.2020.01.003 [34] Y. F. Hernández-Julio, M. J. Prieto-Guevara, W. Nieto-Bernal, C. Jiménez-Velásquez, and J. Ruíz-Guzmán, "Fuzzy knowledge discovery and decision-making through clustering and dynamic tables: Application to bioengineering," in FSDM, 2018, pp. 480-487. [35] Y. F. Hernández-Julio, M. J. Prieto-Guevara, and W. Nieto-Bernal, "Fuzzy clustering and dynamic tables for knowledge discovery and decision-making: Analysis of the reproductive performance of the marine copepod Cyclopina sp," Aquaculture, p. 735183, 2020/02/29/ 2020. https://doi.org/10.1016/j.aquaculture.2020.735183 [36] Y. F. Hernández-Julio, I. Meriño-Fuentes, R. R. González-Díaz, A. Guerrero-Avendaño, L. V. O. Toledo, and W. N. Bernal, "Fuzzy knowledge discovery and decision-making through clustering and Dynamic tables: Application in Colombian business finance," in 2020 15th Iberian Conference on Information Systems and Technologies (CISTI), 24-27 June 2020 2020, pp. 1-5. Appendix A https://drive.google.com/file/d/15dgUc58aOSaPRSs-dHZw9Qn_natgU-E8/view?usp=sharing (accessed on 19 March 2024). Appendix B https://drive.google.com/file/d/1oTTVn5a5lClmh5oKm6ClI5TuVIsG73cf/view?usp=sharing (accessed on 19 March 2024). Declaration of Generative AI and AI-assisted technologies in the writing process While preparing this article, the authors used ChatGPT premium to improve their readability and language because the authors are not native English speakers. After using this tool, the authors reviewed and edited the content as needed and took full responsibility for the content of the publication. https://doi.org/10.1007/978-3-662-43505-2_14 https://doi.org/10.1007/s40815-023-01665-0 https://doi.org/10.1002/9781394214211.ch3 https://doi.org/10.1016/j.eswa.2015.05.006 https://doi.org/10.1016/j.bbe.2019.05.006 https://doi.org/10.1016/j.bbe.2020.01.003 https://doi.org/10.1016/j.aquaculture.2020.735183