To develop and validate an AI model based on deep learning to differentiate between healthy voices, spasmodic dysphonia (SD), and muscle tension dysphonia (MTD) using voice audio recordings.
Approach:
Data Collection: A retrospective analysis of 1,597 voice samples (595 healthy, 471 MTD, 531 SD) was conducted.
Audio Processing: Voice audio was processed into Log-Mel spectrograms.
Model Development: Pre-trained convolutional neural networks (CNNs) including VGG16, ResNet50, and DenseNet161 were used for transfer learning.
Performance Evaluation: The model's performance was evaluated on a held-out test set and compared to the diagnostic assessments of four otolaryngologists.
Key Findings:
The AI model achieved an accuracy of 89.5% in distinguishing healthy voices from disordered ones.
For ternary classification, the model attained an accuracy of 71.6%, with class-specific AUCs of 0.957 (Healthy), 0.731 (MTD), and 0.855 (SD).
The AI model outperformed human experts, who achieved average accuracies of 78.2% in binary classification and 60.6% in ternary classification.
Interpretation:
The deep learning model demonstrates favorable performance in distinguishing SD from MTD using only voice audio signals, comparable to experienced clinical specialists.
Limitations:
The study is based on a specific dataset of Mandarin pathological voices, which may limit generalizability.
The model's performance in real-world clinical settings needs further validation.
Conclusion:
The technology serves as a promising objective auxiliary tool for the preliminary screening and differential diagnosis of voice disorders.
