Machine Learning-Based Cough Tone Classification: Diagnostic Exploration of Chronic Obstructive Pulmonary Disease and Respiratory Tract Infections

This research focuses on acoustic diagnostics, utilizing voice signal analysis and machine learning to extract distinctive voice features for non-invasive early classification of chronic obstructive pulmonary disease and respiratory tract infections. The recent developments in this field include AI-driven voice analysis, machine learning techniques such as convolutional neural networks and support vector machines, signal processing tools like MFCCs, and variable acoustic sensors for detecting disease-related patterns in sound signals. One of the main challenges in the clinical translation of voice-based diagnostic is data scarcity.

Other challenges include limited model generalization, privacy ethics, conflicts, and barriers to interpretability. After assembling the vocal feature indicator database, open SPSS and load the appropriate data file. From the menu bar, select Analyze, then choose Nonparametric Tests, followed by Legacy Dialogues, and click on 2 Independent Samples.

In the pop-up dialogue box, select the observed variables to be compared under the Test Variable List section. Then, under Grouping Variable, select the variable that will be used for grouping. Click the Define Groups button and enter the identifiers for the two groups in the pop-up window.

Under Test Type, select the Mann-Whitney U test. Click OK to run the test and allow SPSS to automatically generate the output. For principal component analysis, ensure that the data is collated, saved in Excel or CSV format, and imported into SPSS version 20.0.

To open the file, select File, then choose Open, followed by Data, and select the appropriate file. To initiate principal component analysis, click Analyze, then choose Dimension Reduction, and select Factor. In the dialogue box, add all continuous variables used in principal component analysis into the Variables field.

Click the Extraction button and select the Principal components method as the extraction technique. Select Eigenvalues greater than 1 as the criterion for retaining principal components. Select the rotation method and click Rotation to choose either Varimax or Promax.

Under Options, check both Scree plot and Coefficient matrix to include the gravel diagram and the matrix of coefficients in the output for evaluating retained variants. After completing all the settings, click OK to execute the analysis and allow SPSS to generate the output. Interpret the principal component loading matrix to assess the relationship between the principal components and the original variables.

Identify variables with higher loading values, as these contribute more significantly to component changes. Use the Total Variance Explained table to evaluate how much variance each principal component accounts for. Identify the principal components with large variance proportions, as they typically capture most of the data variation.

Refer to the scree plot to determine which components to retain. Locate the inflection point and keep all components to the left of this point. If principal component scores are required, check Save as variables before running the analysis.

SPSS will add the scores for each sample as new variables in the dataset. Principal component analysis identified six major components which together accounted for 76.8%of the total variance. The logistic regression model demonstrated stable performance across three validation folds, with AUC values of 0.71, 0.74, and 0.88, yielding a mean AUC of 0.77.

In contrast, the random forest model exhibited greater variability, with fold AUC scores of 0.69, 0.52, and 0.83, and a lower mean AUC of 0.68. The logistic regression model achieved 100%correct predictions for COPD and six out of seven correct for respiratory tract infections, as shown in the confusion matrix, indicating high classification accuracy. The random forest model misclassified one COPD and two respiratory tract infection cases, resulting in lower classification accuracy compared to the logistic regression model.

On the test dataset, the logistic regression model yielded excellent classification performance, achieving an AUC value of 0.95. The random forest model showed lower test performance with an AUC value of 0.76.