Machine Learning-Based Estimation of Hoarseness Severity Using Acoustic Signals Recorded During High-Speed...

Area of Science:

• Laryngology
• Speech Pathology
• Biomedical Engineering
• Machine Learning

Background:

• Assessing hoarseness severity is crucial for diagnosing voice disorders.
• Machine learning offers a promising avenue for objective voice analysis.
• High-speed videoendoscopy (HSV) provides detailed laryngeal visualization but its acoustic recordings' utility for hoarseness assessment is under investigation.

Purpose of the Study:

• To investigate the efficacy of sustained phonations recorded during HSV for machine learning-based hoarseness severity assessment.
• To compare the performance of HSV-derived acoustic recordings with conventional recordings from voice therapy sessions.
• To identify key differences and limitations of HSV-derived acoustic data for voice analysis.

Main Methods:

• A database of 617 voice recordings (250 ms) from HSV examinations was created.
• Comparison databases included 809 vowels from voice therapy sessions (1-second and 250 ms durations).
• Extracted 490 acoustic features, developed machine learning models, and classified hoarseness severity based on expert auditory-perceptual ratings.

Main Results:

• Logistic regression models achieved classification accuracies of 0.863 (VT-1), 0.847 (VT-2), and 0.742 (HS).
• Correlation between predicted and subjective hoarseness scores was 0.797 (VT-1), 0.763 (VT-2), and 0.637 (HS).
• Correlation between changes in quantitative and subjective ratings was significantly lower for HSV recordings (0.088) compared to voice therapy recordings.

Conclusions:

• Acoustic signals from HSV show potential for quantitative hoarseness assessment but are less reliable than voice therapy recordings.
• Practical challenges during oral laryngeal examination limit the quality of HSV-derived acoustic recordings.
• Future improvements, potentially using flexible nasal endoscopy, could enhance the utility of HSV for voice assessment.