Wearable systems play an important role in continuous health monitoring and can contribute to early detection of abnormal health-related events and facilitate the advancement of personalized healthcare. The neck is a unique sensing location because it provides access to a set of health-related data that other wearable devices simply cannot obtain. Activities including breathing, chewing, clearing the throat, coughing, swallowing, speech and even heartbeat can be recorded from around the neck. Two applications of particular interest for this project include medication adherence monitoring and food intake monitoring.
Medication non-compliance, especially for patients with chronic illnesses, is a global issue that has been associated with increased healthcare cost, rehospitalization, complications and disease progression. To address this problem, it is essential to have a portable, wearable health platform that can remind patients of their medication regimen, track medication ingestion, and monitor a patient's overall health status. The proposed system in the form of a necklace will automatically track medication ingestion using the well-established radio frequency (RF) technology in very high frequency (13.56MHz) band. For power management purposes, the system will be 'asleep' by default except during a swallowing event when there is a possibility of medication ingestion. For this reason, automatic swallowing detection is essential; the ability to differentiate swallowing sounds from other tracheal sounds initiated by speaking, coughing, clearing the throat etc. In previous work, we developed a real-time swallowing detection algorithm based on acoustic signals and patterns that combines computationally-inexpensive features to achieve comparable performance with previously proposed offline methods using acoustic and non-acoustic data. With data from four healthy subjects that includes common tracheal events such as speech, chewing, coughing, clearing the throat, and swallowing of different liquids, our results show an overall recall performance of 79.9% and precision of 67.6%, which are slightly better or close to the offline results.
In our following work, we expanded our scope and explored tracheal activity recognition using a combination of promising acoustic features from related work and apply simplistic classifiers including K-NN and Naive Bayes. For wearable systems in which low power consumption is of primary concern, we have shown that with a sub-optimal sampling rate of 16 kHz, we achieved average classification results in the range of 86.6% to 87.4% using 1-NN, 3-NN, 5-NN and Naive Bayes. All classifiers obtained the highest recognition rate in the range of 97.2% to 99.4% for speech classification. This is promising to mitigate privacy concerns associated with wearable systems interfering with the user's conversations.
In Georgia Tech Bionics lab (GT-Bionics) we design and develop state-of-the-art medical and scientific instruments for a wide variety of clinical and research applications. More specifically, our focus is on Assistive Technologies, Rehabilitation Engineering, Wearable Devices for Smart Health and Wellbeing, Implantable Microelectronic Devices, and Wireless Neural Interfacing. We are involved in true multidisciplinary research addressing all aspects of complex biomedical systems from hardware, software, and smart algorithm design to evaluation of their full functionality and efficacy on animal subjects or in clinical settings.
