Automatic detection of swallowing events by acoustical means for applications of monitoring of ingestive behavior

Abstract

Our understanding of etiology of obesity and overweight is incomplete due to lack of objective and accurate methods for monitoring of ingestive behavior (MIB) in the free-living population. Our research has shown that frequency of swallowing may serve as a predictor for detecting food intake, differentiating liquids and solids, and estimating ingested mass. This paper proposes and compares two methods of acoustical swallowing detection from sounds contaminated by motion artifacts, speech, and external noise. Methods based on mel-scale Fourier spectrum, wavelet packets, and support vector machines are studied considering the effects of epoch size, level of decomposition, and lagging on classification accuracy. The methodology was tested on a large dataset (64.5 h with a total of 9966 swallows) collected from 20 human subjects with various degrees of adiposity. Average weighted epoch-recognition accuracy for intravisit individual models was 96.8%, which resulted in 84.7% average weighted accuracy in detection of swallowing events. These results suggest high efficiency of the proposed methodology in separation of swallowing sounds from artifacts that originate from respiration, intrinsic speech, head movements, food ingestion, and ambient noise. The recognition accuracy was not related to body mass index, suggesting that the methodology is suitable for obese individuals.

Fig. 1

a) a 4.0s fragment of a sound recording including a swallow; b) features extracted by WPD processing; c) features extracted by msFS processing. Frequencies are shown for the center of each packet or bin.

Fig. 2

Examples of: a) true positive, b) false positive, c) true negative, d) false negative. Each number represents a class label for an epoch (‘−1’ – non-swallow epoch, ‘1’ – swallow epoch).

Fig. 3

Accuracy of swallowing sound recognition as a function of epoch duration and decomposition level a) msFS with no lags b) WPD with no lags c) msFS with K=1 (3 lags) d) WPD with 3 lags.

Fig. 4

Distribution of average weighted accuracy in classification of epochs and swallowing events versus subject’s BMI and corresponding linear fit of the data.