From voice biomarkers to telemedicine screening: developing and evaluating a voice-based AI model for laryngeal lesion detection using the Bridge2AI-Voice dataset - Report - MDSpire
Advertisement
From voice biomarkers to telemedicine screening: developing and evaluating a voice-based AI model for laryngeal lesion detection using the Bridge2AI-Voice dataset
Clinical Report: Advancing Laryngeal Lesion Detection with AI
Overview
This study evaluates a voice-based AI model using the Bridge2AI-Voice dataset for screening laryngeal lesions. The model demonstrated high sensitivity and moderate specificity.
Background
Current screening for laryngeal lesions relies heavily on in-person laryngoscopy, which can be resource-intensive and limit access to care. The human voice contains significant acoustic information that may indicate laryngeal pathology. The Bridge2AI-Voice initiative aims to provide a large, standardized dataset to enhance the development of AI models for this purpose.
Data Highlights
Metric
Value
95% CI
AUC
0.812
0.744–0.876
Sensitivity
0.870
0.767–0.939
Specificity
0.566
0.479–0.651
Key Findings
The AI model achieved a cross-validated AUC of 0.812, indicating good discrimination ability.
Sensitivity was high at 0.870, while specificity was moderate at 0.566.
Model performance significantly exceeded an age-only baseline (DeLong p = 0.0008).
Subgroup analysis showed consistent sensitivity across benign and precancerous lesions.
Alternative feature modalities did not enhance model performance beyond the OpenSMILE features.
Clinical Implications
Further validation in larger cohorts is warranted to confirm these results.
Conclusion
Future research should focus on larger, prospectively recruited cohorts for confirmatory analysis.
