Explainable Machine Learning for Voice Disorder Screening Using Acoustic Features From Multiple Speech Tasks

Area of Science:

• Speech pathology and acoustics
• Machine learning in healthcare
• Explainable Artificial Intelligence (XAI)

Background:

• Distinguishing pathological voices from normal controls is crucial for early diagnosis and treatment.
• Traditional methods rely on subjective perceptual evaluation, which can be inconsistent.
• Objective acoustic analysis offers a promising avenue for voice disorder assessment.

Purpose of the Study:

• To evaluate machine learning models for pathological voice detection using various acoustic feature sets.
• To compare the performance of different machine learning classifiers.
• To apply Explainable AI techniques to identify key acoustic predictors of dysphonia.

Main Methods:

• Analysis of 1111 voice samples using four acoustic feature sets: sustained vowel, connected speech, concatenated, and combined.
• Training and evaluation of Support Vector Machine, Random Forest, and eXtreme Gradient Boosting classifiers.
• Performance assessment using accuracy and Area Under the ROC Curve, with SHapley Additive exPlanations for feature interpretation.

Main Results:

• The combined feature set with eXtreme Gradient Boosting achieved the highest accuracy (97.31%).
• Acoustic features from multiple speech tasks significantly impacted performance.
• Key predictors included CSID, CPP_v, L/H ratio_v, L/H ratio_SD, and AVQI.
• Misclassifications were limited to mild dysphonia grades.

Conclusions:

• Integrating acoustic data from diverse speech tasks enhances pathological voice discrimination.
• Explainable machine learning successfully identified clinically relevant acoustic predictors.
• This approach shows potential as an objective and transparent tool for voice disorder screening.