Voice disorder classification using convolutional neural network based on deep transfer learning.

Area of Science:

• Medical Informatics
• Machine Learning
• Signal Processing

Background:

• Voice disorders are prevalent globally, necessitating accurate diagnostic tools.
• Machine learning models for voice disorder classification require substantial training data, which is often difficult to obtain due to medical data sensitivity.
• Existing methods face challenges with high-dimensional features and potential overfitting.

Purpose of the Study:

• To propose a novel transfer learning framework, OpenL3-SVM, for the automatic recognition of multi-class voice disorders.
• To address the challenge of limited medical data for training machine learning models.
• To improve the accuracy and efficiency of voice disorder classification.

Main Methods:

• A pretrained convolutional neural network (OpenL3) was utilized to extract high-level feature embeddings from voice signal Mel spectrums.
• Linear Local Tangent Space Alignment (LLTSA) was applied for feature dimension reduction to mitigate overfitting.
• A Support Vector Machine (SVM) classifier was trained on the reduced features for multi-class voice disorder classification.
• Fivefold cross-validation was employed to evaluate the framework's performance.

Main Results:

• The OpenL3-SVM framework demonstrated effective automatic classification of multi-class voice disorders.
• Experimental results indicated that OpenL3-SVM outperformed existing methods in classification performance.
• The framework successfully handled feature extraction, dimension reduction, and classification.

Conclusions:

• The proposed OpenL3-SVM transfer learning framework offers a promising solution for automatic voice disorder recognition.
• This approach effectively overcomes the limitations of small sample sizes in medical datasets.
• OpenL3-SVM has the potential to serve as an auxiliary diagnostic tool for physicians in the future.