Research on Pathological Voice Recognition Based on XGBoost

Summary

Here, we present a protocol for establishing a non-invasive XGBoost-based AI model for diagnosing vocal cord polyps using the Saarbrücken database.

Abstract

With the continuous growth of human social communication, the number of people suffering from voice disorders is also increasing. Due to the objective and non-invasive advantages of acoustic detection methods for pathological voice, the use of speech signal analysis for pathological voice recognition has become a research hotspot. This article first selected 101 continuous vowels /a/ from the German SVD database as the research object. Secondly, using wavelet packet technology for time-frequency analysis, four nonlinear dynamic parameters, namely approximate entropy, sample entropy, fuzzy entropy, and permutation entropy, are extracted from the sub signals as the feature parameter set for the pathological voice classifier. Finally, the machine learning algorithm XGBoost is selected as the pattern recognition method to establish a pathological voice classifier, and the classification performance is verified using five fold cross validation and ROC curve. Experimental results have shown that the accuracy of XGBoost's pathological voice classifier is 0.857, the F1 score is 0.875, and the AUC value is 0.944, all of which are higher than the classifier constructed by SVM, indicating that XGBoost has better performance in pathological voice recognition.

Introduction

The number of people suffering from voice disorders is constantly increasing. According to statistics, one-third of the population experiences voice problems at some point in their lives¹. For professional voice users such as singers and teachers, due to their frequent misuse and abuse of their voice, the incidence of throat diseases is higher. Two studies conducted surveys on teachers in Tianjin and Urumqi, and the results showed that the probability of suffering from voice disorders is 33.81% and 28.23%, respectively^2,3. As a result, there is increasing emphasis on the detection and diagnosis of pathological voice in clinical practice. At present, subjective evaluation methods, laryngoscopy examination, and acoustic detection methods are mainly used for the diagnosis of voice diseases in clinical practice^4,5. Acoustic detection method converts voice signals into digital signals for research, ensuring objectivity and avoiding patient pain during the examination⁶. Therefore, acoustic detection has become an effective auxiliary tool for doctors in clinical diagnosis, and research on it is increasing.

The most important techniques for diagnosing pathological voice using acoustic analysis methods are feature extraction and pattern recognition. Since the 1960s, researchers have been extracting some traditional acoustic parameters for the study of pathological voice, and have found many effective and robust acoustic feature parameters, such as fundamental frequency perturbation, amplitude perturbation, and Mel-frequency cepstral coefficients (MFCC)⁷. In recent years, researchers have not only limited themselves to studying traditional acoustic characteristic parameters, but nonlinear dynamic parameters have also begun to be used in the field of pathological voice recognition⁸. It also has a very good effect. With the rapid development of machine learning, more pattern recognition methods with fast running speed and good classification performance have emerged. There are also more and more pattern recognition methods used in pathological voice recognition, such as support vector machines (SVM), random forest, neural networks, deep learning^9,10. XGBoost is a pattern recognition method that has gained attention in recent years¹¹. It does not require feature normalization and can select features on its own. It can adapt to various loss functions and uses regularization terms to prevent overfitting. It has the characteristics of fast speed, portability, and fault tolerance. Therefore, this article attempts to apply the XGBoost method to pathological voice recognition in order to achieve better recognition results.

Protocol

The technical workflow of this study is illustrated in Figure 1.

1. Acquisition of voice samples

1. Source the experimental data from the SVD in Germany¹².
2. Obtain 101 cases of vowel /a/ voice data, including 45 cases (19 males and 26 females) of vocal polyp patients and 56 cases (28 males and 28 females) of normal individuals.

2. Wavelet packet decomposition of voice data¹³

1. Conduct an exhaustive search across three Daubechies wavelets (db1, 3, 5) and three decomposition depths (4, 6, 8 levels) using software MATLAB R2023a.
2. Use a 16 kHz sampling rate, a four-level db3 wavelet packet decomposition (WPD) to partition the speech signal into 16 sub-bands (500 Hz resolution each) (Figure 2).

3. Feature parameter extraction

1. Extract discrete coefficient sequences from the 16 sub-bands obtained via 4-level WPD by Python 3.10, serving as the input values for all entropy equations described below.
   NOTE: All equations used in this study were derived from previously published algorithms (as in references) and were not independently derived.
2. Calculate the approximate entropy for the 16 WPD subband sequences using the formula below¹⁴:
       (1)
       (2)
   Cim (r)={d[X(i), X(j)]＜r} / (N-m+1)     (3)
   Parameters: The pattern dimension m is chosen as 2, and the similarity tolerance r is selected as 0.1 to 0.25 times the standard deviation¹⁴.
3. Calculate the sample entropy for the 16 WPD subband sequences using the formula below¹⁵:
        (4)
       (5)
   B_i^m (r) = {d[X(i), X(j)]＜r} / (N-m)    (6)
   Parameters: The pattern dimension m is chosen as 1 or 2, and the similarity tolerance r is selected as 0.1 to 0.25 times the standard deviation¹⁵.
4. Calculate fuzzy membership function with the formula below¹⁶:
        (7)
5. Calculate the fuzzy entropy with the formula below¹⁷:
       (8)
       (9)
       (10)
   Parameters: The pattern dimension m is chosen as 2, and the similarity tolerance r is selected as 0.15 times the standard deviation.
6. Calculate the permutation entropy with the formula below¹⁸:
        (11)
   Parameters: The pattern dimension m is chosen as 6, and time delay L = 2 was ultimately determined for permutation entropy extraction.

4. Model establishment

1. Divide voice samples from normal and patients with vocal cord polyps into a training dataset and a test dataset at an 8:2 ratio.
2. Use the training dataset for parameter tuning and model training, and then apply the resulting classifier to the test dataset for disease classification.
3. Perform the SVM modeling procedure using Python 3.10.
   1. Confirm the nonlinear nature of the data by comparing SVM kernel performances. Initial tests with a linear kernel yielded poor accuracy, whereas the radial basis function (RBF) kernel significantly improved classification results, demonstrating that the feature space requires a nonlinear decision boundary.
   2. Use the 16-dimensional entropy feature vectors (extracted via WPD) as input for the SVM classifier. Implement an RBF kernel to map the nonlinear voice features into a high-dimensional space.
   3. Perform a grid search using Python (scikit-learn) to identify the optimal combination of the penalty coefficient C and the kernel width γ during the training phase. Evaluate each parameter set by maximizing the cross-validation recognition rate.
   4. Train the SVM model using voice samples from normal individuals and patients with vocal cord polyps. Use the optimal hyperparameters obtained from the grid search to construct the final trained model.
   5. Apply the trained model to unseen test samples. Each test sample undergoes the same WPD and entropy extraction procedure as the training data. The SVM predicts the binary class label (normal vs. vocal cord polyps) based on the spatial position of the test feature vector relative to the optimized decision boundary.
4. Perform XGBoost modeling procedure using Python 3.10.
   1. Use a Python-based grid search to optimize key hyperparameters (including learning rate and tree depth) during the training phase. Evaluate multiple parameter combinations via cross-validation to identify the configuration that maximizes classification accuracy.
   2. Train a binary XGBoost classifier using the sixteen entropy features extracted from voice samples. Apply the optimal hyperparameters obtained from the grid search to construct the final model.
   3. Test the trained model on an independent validation set consisting of unseen samples. This step assesses the classifier's generalization capability by evaluating its performance on data not used during training.

5. Evaluation of the model performance

1. Conduct a five-fold cross-validation method.
   1. Partition the preprocessed voice dataset randomly into five equal folds. Use four folds as the training set to construct the model and use the remaining one fold as the validation set for performance assessment.
   2. Repeat this process five times. Use the average of the results from the five iterations as the final model performance.
2. Obtain the confusion matrix.
   1. Generate the confusion matrix by comparing the classifier's predicted labels with the ground-truth labels for all test samples according to the five-fold cross-validation results. Aggregate these individual comparisons into a contingency table using the Python scikit-learn library to quantify correct and incorrect classifications, as shown in Table 1.
3. Calculate accuracy, precision, sensitivity, and F1 score to describe the effectiveness of a classification model. The relevant calculation formulas are as follows.
   Accuracy = (TP + TN)/(TP + TN + FP + FN)
   Precision = TP/(TP + FP)
   Sensitivity = Recall = TPR = TP/(TP+ FN)
   F1 = (2 × Precision × Recall)/ (Precision + recall)
   NOTE: These metrics (TP, TN, FP, FN) were derived from the elements of the confusion matrix.
4. Describe the ROC curve with AUC.
   1. Plot the ROC curve to visualize the trade-off between the true positive rate (TPR) and the false positive rate (FPR) across all classification thresholds. Calculate the Area Under the Curve (AUC) as a summary metric to quantify the model's overall discriminative ability.
      NOTE: An AUC value closer to 1 indicates a higher probability that the classifier correctly distinguishes between normal individuals and those with vocal cord polyps, independent of the specific decision threshold.

Results

In this study, wavelet packet analysis was employed to perform time-frequency decomposition of voice signals. By integrating nonlinear dynamic parameters—including approximate entropy, sample entropy, fuzzy entropy, and permutation entropy—the complexity and irregularity characteristics of pathological voices associated with vocal fold polyps were systematically characterized. Two pathological voice classifiers were subsequently constructed based on the support vector machine (SVM) algorithm and the XGBoost algorithm, re...

Discussion

Pathological voice diagnosis utilizing speech signal analysis represents a non-invasive and objective approach¹⁹. Consequently, pathological voice classification has emerged as a significant research focus in speech signal processing and recognition, demonstrating substantial clinical application value and developmental prospects²⁰. This study employed speech samples collected from SVD as experimental corpora. Through speech signal processing and machine learning techniques...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
MATLAB	The MathWorks	R2023a	Primary software platform for numerical computation and algorithm prototyping.
Python software	Python Software Foundation	Python 3.10	Core programming language used throughout the study.
Saarbruecken Voice Database,SVD	Institute of Phonetics, Saarland University	Saarbrücken Voice Database (SVD)	Publicly available database of pathological and normal voice recordings. Contains sustained vowels and continuous speech from subjects with conditions including vocal cord polyps, as well as healthy controls. Used for academic research under its specified terms of access.