Pathological voice classification based on multi-domain features and deep hierarchical extreme learning machine.

Area of Science:

• Medical Informatics
• Signal Processing
• Machine Learning

Background:

• Pathological voice signal screening is crucial for computer-aided diagnosis.
• Non-invasive, real-time diagnostic tools are increasingly important for clinicians and researchers.

Purpose of the Study:

• To propose a novel method for automatic identification of voice disorders using multi-domain features and a hierarchical extreme learning machine (H-ELM).
• To enhance the efficiency and reduce overfitting in pathological voice classification.

Main Methods:

• Extraction of sensitive features from voice signals using multi-domain analysis (time domain, sample entropy, ensemble empirical mode decomposition, gammatone frequency cepstral coefficients).
• Application of neighborhood component analysis for high-dimensional feature reduction and selection.
• Training a hierarchical extreme learning machine (H-ELM) with selected sensitive features for classification.

Main Results:

• The proposed H-ELM achieved high performance metrics: 99.37% sensitivity, 98.61% specificity, 99.37% F1 score, and 98.99% accuracy.
• The method demonstrated effectiveness in filtering sensitive features and improving classification efficiency.

Conclusions:

• The developed multi-domain feature extraction and H-ELM approach is a feasible and highly accurate method for the initial classification of pathological voice signals.
• This intelligent data-driven screening tool shows significant potential for computer-aided diagnosis of voice disorders.