Pathological Voice Detection Based on Phase Reconstitution and Convolutional Neural Network.

Area of Science:

• Acoustic analysis
• Biomedical signal processing
• Machine learning for healthcare

Background:

• Nonlinear dynamic features effectively characterize normal and pathological voice acoustics.
• Accurate classification of voice disorders is crucial for diagnosis and treatment.
• Limited clinical data often hinders the development of robust voice analysis models.

Purpose of the Study:

• To classify normal and pathological voice using phase space reconstruction and convolutional neural networks.
• To address the challenge of limited clinical data through a data enhancement scheme.
• To evaluate the proposed method's performance and robustness on diverse voice databases.

Main Methods:

• Phase space reconstruction to convert 1D voice signals into 2D trajectory graph samples using delay time and embedding dimension.
• VGG-like convolutional neural network (CNN) architecture to extract graphical features from trajectory graph samples.
• Data enhancement techniques to augment limited clinical datasets.
• Five-fold cross-validation for performance evaluation on the MEEI, SVD, and a custom clinical database.

Main Results:

• Achieved average recognition rates of 99.42% (MEEI), 97.30% (SVD), and 95.88% on the custom database.
• Demonstrated high accuracy in distinguishing normal, vocal fold paralysis, and vocal fold non-paralysis voices (96.04% and 92.27% on MEEI and SVD, respectively).
• The method exhibited high classification accuracy, good robustness, and broad applicability for voice disorder recognition.

Conclusions:

• The combination of phase space reconstruction and CNN provides an effective approach for classifying normal and pathological voices.
• The data enhancement scheme successfully mitigates issues related to insufficient clinical data.
• The proposed method shows significant potential for widespread application in voice disorder diagnosis.