Linear versus nonlinear modeling of dysphonia severity: A comparison between Acoustic Voice Quality Index and machine learning

Abstract

Purpose: This study compares the Acoustic Voice Quality Index (AVQI-3) with machine learning (ML) models to evaluate their clinical utility for estimating voice quality. Methods: Audio from 187 American English speakers (49 healthy, 138 with voice disorders) was rated for overall voice quality by six voice specialists. AVQI-3 and its six acoustic parameters were extracted, and these parameters were used to train 14 ML models (linear, curvilinear, nonlinear). Correlations and classification accuracy (normal–mild vs. moderate–severe) were compared against perceptual ratings as ground truth. Results: AVQI-3 correlated strongly with perceptual ratings (Spearman rs=0.75), comparable to linear and curvilinear models, yet outperformed most nonlinear models. At a cutoff score of 2.52, AVQI-3 achieved the highest classification accuracy (0.92) with balanced sensitivity (0.90) and specificity (0.93). Among ML models, linear regression performed best (rs=0.77, accuracy=0.92, sensitivity=1.0, specificity=0.89), whereas nonlinear models showed reduced performance (average rs=0.74, accuracy=0.87, sensitivity=0.95, specificity=0.85). Conclusion: AVQI-3 is a simple, accessible index that quantifies voice quality as effectively as complex ML models. This is supported by the best-performing ML models being linear, indicating that linear combinations of acoustic measures are effective, accurate, and clinically interpretable, whereas added nonlinear complexity offered little benefit.