Replication-based regularization approaches to diagnose Reinke's edema by using voice recordings.

Area of Science:

• Medical acoustics
• Speech pathology
• Machine learning

Background:

• Reinke's edema is a common laryngeal pathology.
• Computer-aided diagnosis (CAD) systems can detect Reinke's edema using speech features.
• Within-subject variability in speech recordings necessitates specialized statistical methods.

Purpose of the Study:

• To develop and evaluate novel regularization-based approaches for classifying Reinke's edema.
• To specifically address and incorporate within-subject variability in statistical models for laryngeal pathology detection.
• To compare the performance of replication-based regularization methods against traditional independence-based approaches.

Main Methods:

• Extraction of acoustic features from four phonations of the sustained vowel /a/ for 30 Reinke's edema patients and 30 healthy subjects.
• Implementation of three replication-based regularization approaches for variable selection and classification.
• Utilizing a cross-validation framework to assess the reliability and predictive ability of the proposed methods.
• Comparison with traditional independence-based regularization methods.

Main Results:

• The proposed replication-based approaches demonstrated reliability in feature selection and predictive performance.
• A stable accuracy rate of 0.89 was achieved under a cross-validation framework.
• Traditional independence-based methods exhibited significant variability in selected features and accuracy metrics.
• The novel methods effectively addressed within-subject variability.

Conclusions:

• The developed replication-based regularization approaches are reliable for detecting Reinke's edema.
• These methods offer a robust solution for statistical analysis in the presence of within-subject variability.
• The findings contribute to the development of more accurate expert systems for laryngeal pathology diagnosis.