Applying Machine Learning Algorithms for Automatic Detection of Swallowing from Sound.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Dysfunctional swallowing has severe consequences, including aspiration pneumonia and reduced quality of life.
• Current methods for monitoring swallowing are limited to clinical settings, lacking simple, non-invasive solutions.
• Accurate swallowing detection is crucial for both clinical management and research.

Purpose of the Study:

• To develop a highly accurate, non-invasive method for binary swallowing detection using laryngeal sounds.
• To evaluate the efficacy of supervised machine learning algorithms for analyzing swallowing sounds.
• To explore the potential of laryngeal sound analysis as an accessible swallowing assessment tool.

Main Methods:

• Collected a dataset of 2500 swallow sound samples and 1700 laryngeal noise samples from 15 healthy adults.
• Trained and tested three supervised machine learning algorithms: decision tree, support vector machine (SVM), and neural network (scaled conjugate gradient - SCG).
• Utilized sound recordings for feature extraction and classification of swallowing events.

Main Results:

• The decision tree and neural network (SCG) models achieved an area under the ROC curve of 0.970 and 0.971, respectively.
• Average accuracies for the models were high: 93.2% for the decision tree, 86.2% for SVM, and 93.7% for the neural network (SCG).
• These results demonstrate significant potential for machine learning in accurately detecting swallowing from laryngeal sounds.

Conclusions:

• Machine learning strategies show considerable promise for improving the accuracy of swallowing detection.
• Non-invasive laryngeal sound analysis offers a viable approach for swallowing monitoring outside clinical environments.
• Further optimization and validation are necessary, but initial findings support the development of this technology for clinical and research applications.