| Obesity is a global epidemic which imposes a financial burden and increased risk for a myriad of co-morbidities including diabetes, hypertension, and even death. This dissertation presents the design and implementation of a prototype for an Automated Ingestion Detection process in conjunction with a remote Health Monitoring System (AID-HMS), intended to support existing obesity and overweight management therapies by detecting periods of ingestion from throat sounds. The AID-HMS prototype presented is envisioned to enhance methods of self-reporting of meals, where possibly forgotten periods of ingestion may be identified, as well as offer an estimate of meal timing in both duration and onset.;Ingestion sounds from seven individuals consuming liquids and solids ad libitum are recorded as well as sources of non-ingestion noise such as coughing, clearing the throat, movement, and voice. Ten time and frequency domain features, used in related works and in speech processing, are examined as well as eight cost functions implemented on wavelet based decompositions, from fifty-one distinct wavelets, to determine a combination of feature, feature creation parameters, and machine learning based classifier that offers a high level of ingestion detection accuracy.;Features and classifiers are combined to form Multiple Classifier System (MCS) groups to improve the detection accuracy. Two methods of training set classification are examined to approach the ingestion detection problem from a swallow sound and ingestion sound detection perspective. The result is a "spread" approach, where an estimate of ingestion activity level from three different MCS groups is stored on a remote database, and is viewable through a web browser.;From ten-fold cross-validation, performed on the dataset of recordings, top performing MCS groups achieved swallow detection accuracy of approximately 90% and false positive rate (non-swallow sounds detected as swallows) of 10%. When examining separation between swallow sounds and voice alone, swallow detection accuracies near 98% and false positive rates of 8.5% have been observed.;The flow from sound recording, processing, and remote storage/viewing has been implemented. Future directions will include implementation of the AID process on a small, portable platform as well as performance evaluated on a larger dataset of individuals. |
