Improving Non-Invasive Aspiration Detection With Auxiliary Classifier Wasserstein Generative Adversarial Networks

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Aspiration is a critical complication in swallowing disorders, necessitating accurate detection for effective dysphagia management.
• High-resolution cervical auscultation shows promise as a noninvasive screening tool for swallowing, but its diagnostic algorithms require extensive training data.
• Collecting cervical auscultation data is challenging due to clinical costs, time, and the need for expert interpretation, with rare severe aspiration events limiting machine learning performance.

Purpose of the Study:

• To develop a data augmentation technique using auxiliary classifier Wasserstein generative adversarial networks (AC-WGANs) to generate supplementary training data for cervical auscultation signals.
• To enhance the performance of machine learning models for automatic aspiration detection in dysphagia care.
• To overcome the limitations of insufficient and imbalanced data in training cervical auscultation-based diagnostic algorithms.

Main Methods:

• Utilized auxiliary classifier Wasserstein generative adversarial networks (AC-WGANs) to capture the distribution of original cervical auscultation signal features and generate synthetic data.
• Conducted a 10-fold subject cross-validation on a dataset comprising 2079 sets of 36-dimensional signal features from 189 patients.
• Compared the performance of the proposed data augmentation method against basic data sampling, cost-sensitive learning, and other generative models.

Main Results:

• The AC-WGAN-based data augmentation significantly improved classification performance compared to baseline methods.
• The proposed method demonstrated superior results over basic data sampling, cost-sensitive learning, and other generative models.
• Achieved significant enhancement in the accuracy of classifying cervical auscultation signals for aspiration detection.

Conclusions:

• The developed AC-WGAN approach effectively generates supplementary training exemplars, addressing data scarcity in cervical auscultation analysis.
• This method shows remarkable potential for improving the classification accuracy of noninvasive swallowing evaluation tools.
• Paves the way for developing more accurate and reliable noninvasive dysphagia care solutions through enhanced machine learning models.