Sex-Specific Acoustic Biomarkers of Dysphonia: A High-Resolution ROC Analysis Using the Saarbrücken Voice Database

Dysphonia significantly alters voice quality, necessitating objective and cost-effective clinical assessment tools. However, acoustic metrics often vary based on preprocessing pipelines and dataset composition. This study aims to provide a robust characterization of healthy and dysphonic voices using a highly reproducible signal-processing framework applied to a large-scale database.

Acoustic features (F0, RAP, APQ, NHR, and CPP) were extracted from a standardized 0.4-second stable segment of sustained vowels using a high sampling rate (50 kHz). Group differences were assessed through non-parametric statistics and effect sizes. Diagnostic discrimination was quantified via sex-stratified ROC analysis, including 95% confidence intervals estimated by bootstrapping.

Significant sex-specific patterns emerged: F0 provided moderate discrimination in females (AUC = 0.688), while CPP emerged as the most robust biomarker in males (AUC = 0.744). Perturbation (RAP, APQ) and noise (NHR) metrics showed consistent elevations in dysphonic groups, yielding stable diagnostic cutoffs (eg, RAP: 0.27-0.33%; NHR: 0.007-0.010) with good specificity. The absence of zero-padding and high-resolution sampling improved measurement stability across all clinical subtypes.

Standardized preprocessing without temporal interpolation enhances the reliability of acoustic biomarkers. The established sex-specific normative values and objective diagnostic thresholds provide a robust benchmark for automated voice screening and clinical monitoring, effectively addressing the inherent heterogeneity of clinical databases.