Machine learning based assessment of hoarseness severity: a multi-sensor approach centered on high-speed...

Area of Science:

• Otolaryngology
• Speech-Language Pathology
• Biomedical Engineering

Background:

• Functional voice disorders lack objective assessment due to reliance on subjective evaluations with high inter-rater variability.
• High-speed videoendoscopy (HSV) offers objective insights into vocal fold dynamics, complementing acoustic and clinical data.
• Synchronous recording of HSV and acoustic data enables a unified, objective assessment of voice function.

Purpose of the Study:

• To investigate a machine learning approach for objective hoarseness severity assessment.
• To evaluate the diagnostic value of combining high-speed videoendoscopy (HSV), acoustic, and clinical data.
• To compare the performance of individual and combined data modalities in assessing vocal function.

Main Methods:

• Analysis of 457 HSV recordings, 634 synchronized acoustic recordings, and clinical data from 923 visits.
• Development of a videoendoscopic model using machine learning for hoarseness classification and severity rating.
• Comparison of videoendoscopic, acoustic, and clinical models, including ensemble models combining modalities.

Main Results:

• Individual models showed moderate correlations with subjective hoarseness ratings (videoendoscopic: 0.464, acoustic: 0.512, clinical: 0.638).
• An ensemble model integrating HSV and acoustic data improved correlation to 0.603.
• A multimodal ensemble model incorporating all data types achieved the highest correlation (0.752) with subjective ratings.

Conclusions:

• Synchronous HSV and acoustic recordings offer a more comprehensive and objective assessment of vocal function.
• Multimodal data integration significantly enhances diagnostic accuracy for hoarseness severity.
• Future work should focus on flexible endoscopy and refined parameter extraction for improved robustness.