Mechano-acoustic sensing of physiological processes and body motions via a soft wireless device placed at the suprasternal notch

Mechano-acoustic sensing of physiological processes and body motions via a soft wireless device placed at the suprasternal notch

The human body consists of complex and dynamic systems that network throughout the body. The neck is an essential crossroads of several critical systems under soft tissue in a spatially confined area. These unique characteristics of the neck allow access to a rich set of mechanical and acoustic signals of key dynamic systems – all observable at a single location. In our recent work, recently published in Nature Biomedical Engineering, we demonstrate the application of a novel wireless flexible sensor that can capture these mechano-acoustic signals effectively: Lee, K., Ni, X., Lee, J.Y. et al. Mechano-acoustic sensing of physiological processes and body motions via a soft wireless device placed at the suprasternal notch. Nat Biomed Eng (2019) doi:10.1038/s41551-019-0480-6.

The sensor design combines low modulus, elastic mechanics with mechanically isolated islands of conventional rigid electronic components along with added reinforcements. Together, the sensor can conform to intricate and irregular mounting surfaces and maintain contact during the natural motion of the neck without causing skin irritation and mechanical failure of the internal electronics.

The sensor is placed on the suprasternal notch (between the jugular notches). This unique mounting location uses the trachea as a grounding point to measure the signals from nearby carotid arteries and the esophagus along with chest wall motion. Since the sensor straddles the manubrium, the independent motion of the chest wall compared to the trachea results in a more apparent respiratory signal. The sensor records 3-axis acceleration at a sampling rate of 1600 Hz for the direction normal to the skin, and 200 Hz for  in-plane movements along the surface of the skin. With this wide frequency band, the sensor can capture both high-frequency vibrations, such as vocalizations, heart sounds, and swallowing sounds, as well as low-frequency motion, such as respiration, heartbeats, laryngeal movements, body movements, and body orientation.

The captured physiological information and the system noise superpose onto each other in the measured acceleration signal. To parse out the rich set of underlying physiological information, we first assessed the sensor noise level (a). Power spectral density of the z-axis accelerations collected from a device vertically resting on an elastomer is compared to that of a device interfacing to the suprasternal notch of a subject sitting quietly. The results show that the system noise level is about 10-3 g/√Hz. The comparison also highlights that various homeostatic physiological events induce measurable signals in the 0 – 100 Hz range and have a significantly higher power level than the system noise. We then gathered multiple non-periodic voluntary acceleration signals to study the unique characteristics and features of physiological signals (b). Each of the signals, whether voluntary or involuntary, show distinct frequency and time-domain characteristics (c). Visualizing both time domain and frequency domain information via spectrogram analysis, we see that the heart activity signals are composed of periodic 20 to 50 Hz bursts. For a talking signal, the spectrogram features a harmonic series across a wide band. The swallowing signal comprises sequential events, including low-frequency laryngeal movement followed by the wide-band sound of food going down the esophagus. The walking signal has relatively large amplitude, low-frequency accelerations. Using these features, we developed an algorithm that derives heart rate, respiration rate, energy expenditure, swallow counts, and talk time.

For now, there remains a challenge in deriving the information in an ambulatory setting where different activity signals may overlap. Specifically, when the unrelated movement shares the same frequency band or is in phase with the desired signal, it is difficult to extract unambiguous information. Therefore, we chose to begin our field studies in a relatively quiet environment with minimal movement: sleep. Having the ability to noninvasively monitor physiological signals with a small, conformal sensor that does not interfere with physiological processes or behaviors addresses major issues with conventional sleep monitoring systems. Particularly, without the interference of complex network of wires and uncomfortable sensors used in the conventional sleep monitoring system (i.e. polysomnogram (PSG)), the sleep environment in the clinic will more closely resemble that of typical nights at home, and allow for more accurate assessments of sleep disorders without the constraints of location. With the relatively clean signal captured by our device in this setting, we can accurately extract the essential physiological information (e.g. heart rate, respiration, and body orientation), and detect clinical events (e.g. sleep apnea, restless sleep, and snoring).

Currently, sleep scoring relies on multichannel information acquired from PSG, the gold standard for clinical sleep studies. As the acceleration data from our sensor captures physiological signals related to a subset of PSG signals in addition to information such as body orientation, we attempted to derive an estimate of sleep stages using machine learning. Using a Hidden Markov Model, we were able to derive the sleep stages from the MA signal with 82% accuracy in binary wake/asleep classification.

In conclusion, this work shows the scope of the mechanoacoustic sensor’s functionality. Feature extraction currently requires adjustments for noisy ambulatory environments. We have begun partnerships with physicians and researchers hoping to generate a rich dataset of physiological information in healthy and pathological states including sleep neurology, stroke, cerebral palsy, dysphagia, and surgical recovery. With more data from these collaborations, we will be able to build models that automatically classify physiological events and predict health-related outcomes. We hope that our work paves the way for a single sensor that can accurately and continuously monitor physiological signals while conforming to the nonplanar surface of the human body.





Please sign in or register for FREE

If you are a registered user on Nature Portfolio Bioengineering Community, please sign in