(700) A Proposed Deep Learning Architecture for Multi-Class Vocal Pathology Screening Across Common Voice Tasks

Introduction: While recent machine-learning-based vocal screening tools show promise, most models are restricted to binary pathology vs non-pathology classification, limiting their clinical applications. We propose a proof-of-concept, multi-class deep learning model capable of screening for three common vocal pathologies using diverse vocal tasks.

Methods: We utilized mel-spectrograms from the Bridge2AI Voice dataset to develop a multi-task deep learning model with a shared audio feature encoder and separate task-specific classifier heads. This design allows the model to learn both general vocal features and task-specific cues. Model development took place from December 2025-present, using mel-spectrograms collected between January 2024-November 2025 at five academic research centers across the United States. This model focuses on participants with benign vocal fold lesions, muscle tension dysphonia (MTD), unilateral vocal fold paralysis (UVFP), or no pathology (control). Independent variables include vocal tasks of interest, namely sustained phonation, prosodic speech, respiration/cough, and diadochokinetic tests. Outcome metrics include accuracy, sensitivity/specificity, area under the receiver operating characteristic (AUROC), and area under precision recall curve (PR-AUC).

Results: Our final dataset contains 76 control patients (2,401 spectrograms), 50 with benign lesions (1,688 spectrograms), 52 with MTD (1,757 spectrograms), and 61 with UVFP (2,004 spectrograms). In preliminary training, participant-level accuracy achieved 35-45% (chance=25%); participant-level validation and testing AUROCs were 0.62 and 0.65, respectively, with corresponding PR-AUCs of 0.39 and 0.46.

Conclusions: Preliminary results suggest that a task-aware, multi-class vocal screening framework can capture meaningful discriminative signal among common vocal disorders.