Date

2019

Author

Koyuncu, Abdullah Onur

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

117
views

0
downloads

Cite This

Statistics such as swallow frequency, the average time between rhythmic swallows and the maximum number of rhythmic swallows can be related to the feeding maturity of infants. Therefore, detecting swallow segments automatically from an acoustical feeding signal can be considered as a decision support mechanism for neonatologists. This thesis includes different approaches for the analysis of infant's feeding sounds and proposes two different pattern recognition methodologies, segmentation followed by classification and classification followed by merging, for auto-segmentation and classification of swallowing instants. Data from 52 infant subjects are used, in which acoustic feeding signals are recorded with a digital stethoscope. Multiple learning algorithms such as Gaussian mixture models (GMM), support vector machines (SVM) and hidden Markov models (HMM) are used to discriminate swallowing sounds from other sound activities. A comprehensive set of feature extraction methods in time and frequency domain are investigated for the representation of captured acoustic signals. Moreover, feature selection methods are examined thoroughly to improve the representation power of feature vectors. Experimental comparison in terms of precision, recall and F1 scores of eight different paths to segment and classify swallow instants is made. The results show that the first approach segments the swallow episodes with lower performance as the error in the segmentation also affects the classification performance negatively. On the other hand, best results are obtained in the second approach where binary and 3 class SVM classifiers are applied with purpose-specific finite state machine algorithms. In the time duration based performance evaluation, the F1 scores are obtained as almost equal to 0.70 for both methods. On the other hand, they are computed as nearly 0.81 in the event based one.

Subject Keywords

Neonatal intensive care, Neonatal intensive care Technology., Keywords: Classification, Gaussian mixture model, Hidden Markov model, Machine learning, Newborn infants, Pattern recognition, Support vector machine, Swallow sound

URI

http://etd.lib.metu.edu.tr/upload/12623424/index.pdf
https://hdl.handle.net/11511/43632

Collections

Graduate School of Natural and Applied Sciences, Thesis