嚥下音解析による摂取水分量の推定システム

抄録

When humans swallow, swallowing sounds occur at the larynx. Since swallowing sounds can be measured noninvasively, these sounds are expected to be useful for testing swallowing ability. By analyzing these sounds, researchers have attempted to elucidate the mechanisms and the acoustic characteristics of the swallowing sounds. However, little research has been done which utilizes swallowing sounds to measure how much food or liquid is being swallowed. If the amount of food or drink being swallowed can be measured quantitatively, some diseases may be prevented by monitoring eating habits. The purpose of this research was to estimate the volume of water swallowed by analyzing swallowing sounds. First, swallowing sounds were recorded by applying a microphone on the skin surface above the thyroid cartilage when volumes of 5mL, 10mL and 15mL were swallowed by a subject. Fifty measurements were obtained. Second, signal processing methods such as smoothing, linear prediction analysis, Fourier transform and wavelet transform were applied to the swallowing sounds. Eighty-eight feature variables that may reflect the amount of water swallowed were obtained from the results. Using principal component analysis, the number of feature variables was reduced from 88 to 13. Finally, the estimated volume of water swallowed was evaluated using support vector machines. The volume of water was accurately identified with a probability of 72% for volumes between 5mL and 10mL, a probability of 84% for volumes between 10mL and 15mL, and a probability of 88% for volumes between 5mL and 15mL. When the same analysis was performed in several human subjects, the precision of classifying volume of water decreased, which may suggest individual differences in movements of the larynx and pharynx during swallowing.