AI-Driven Detection and Classification of Voice Disorders Using Acoustic Recordings.

Area of Science:

• Medical Artificial Intelligence
• Speech Pathology
• Computational Linguistics

Background:

• Voice disorders affect millions globally, necessitating efficient diagnostic tools.
• Current diagnostic methods can be invasive or resource-intensive.
• AI offers potential for non-invasive, scalable voice disorder assessment.

Purpose of the Study:

• To develop and evaluate AI models for detecting and classifying voice disorders using acoustic recordings.
• To assess the performance of different AI strategies, including HuBERT features and Audio Spectrogram Transformer (AST).
• To explore AI's utility in facilitating earlier diagnosis and optimizing clinical resource allocation.

Main Methods:

• A multicenter study analyzed acoustic recordings from 2613 participants (1948 patients, 665 controls).
• Two AI modeling strategies were employed: HuBERT feature extraction and AST fine-tuning.
• Models were trained and validated using a fixed split with 10-fold cross-validation and an independent test set.

Main Results:

• AI models achieved near-perfect detection of pathological vs. healthy voices (AUROC 0.993).
• Classification of specific voice disorder subtypes showed lower performance (e.g., non-neurological vs. neurological AUROC 0.744).
• HuBERT-based models demonstrated modest performance across various binary classifications.

Conclusions:

• AI, especially AST, excels at distinguishing pathological from healthy voices.
• Improved AI models are required for accurate classification of specific voice disorder subtypes.
• AI-driven acoustic analysis shows promise as a noninvasive screening tool for early voice disorder detection.