Reproducible Machine Learning-Based Voice Pathology Detection: Introducing the Pitch Difference Feature.

Area of Science:

• Speech and Hearing Sciences
• Computational Linguistics
• Biomedical Engineering

Background:

• Voice pathologies affect communication and quality of life.
• Accurate detection of voice disorders is crucial for timely intervention.
• Existing methods may lack robustness or rely on complex feature sets.

Purpose of the Study:

• To develop and validate a novel methodology for voice pathology detection.
• To leverage a publicly available voice database and a comprehensive feature set.
• To explore the efficacy of various machine learning algorithms for this task.

Main Methods:

• Utilized the Saarbrücken Voice Database for analysis.
• Combined established acoustic features with novel pitch difference and NaN features.
• Evaluated six machine learning algorithms (SVM, k-NN, Naive Bayes, Decision Tree, Random Forest, AdaBoost).
• Employed grid search for hyperparameter optimization and extensive feature subset selection.
• Applied k-means synthetic minority oversampling technique to address class imbalance.
• Validated models using repeated stratified cross-validation.

Main Results:

• Achieved high unweighted average recall: 85.61% for females, 84.69% for males, and 85.22% combined.
• Demonstrated the effectiveness of the proposed feature engineering and machine learning approach.
• Omitted accuracy metric due to its bias in imbalanced datasets.

Conclusions:

• The proposed methodology shows significant potential for detecting voice pathologies using machine learning.
• The approach is effective even with a simple vocal task (sustained /a:/ vowel).
• A publicly available GitHub repository and REFORMS checklist are provided to enhance reproducibility and usability.