Classification of Swallowing Foods Using Machine Learning Algorithms

There are limits in assessing healthy and abnormal swallowing by Videofluoroscopic swallowing study. Classification of accelerometric swallowing signals is much more efficient method to judge healthy swallowing. However, these methods have developed mostly with dual axis accelerometric signals and classifying two-class problems. This study is to examine classification methods with multi-class three-axis accelerometric signals. Swallowing signals of five foods are classified with both supervised learning algorithm and unsupervised learning algorithm. Three-axis signals denoised by 10-level discrete wavelet transform with soft thresholding before feature calculation. The result confirmed that classification with support vector machine and K-nearest neighbor can predict with 90% accuracy. However, Classification with fuzzy c-mean clustering produce low purity and normalized mutual information.