Glabella

CONTINUOUSLY SENSING BLOOD PRESSURE BEHAVIOR USING AN UNOBTRUSIVE WEARABLE DEVICE

 
Glabella_banner.png

IMWUT 1, 3, Article 58: Christian Holz and Edward Jay Wang

Microsoft Research, Redmond, WA.


Our wearable glasses prototype Glabella incorporates optical sensors to continuously measure and store the user's pulse waves at three different sites on their face. Our device additionally comprises inertial sensors and a processing unit that compares the continuously recorded pulse waves to extract the user's pulse transit time—the delay between the moments at which the blood ejected from the heart reaches the three sites. Pulse transit time functions as a proxy measurement to monitor the short-term behavior of the user's systolic blood pressure, to which our evaluation shows a significant correlation during in-the-wild use.

2017-imwut17-glabella-evolution_prototype_revisions.jpg

The iterative evolution of our Glabella prototypes. While the first two prototypes consisted of Arduino-based components to study and evaluate the feasibility of our approach in pilots, Versions 3–5 miniaturize the form factor into an integrated device that comprises all components on two separate types of boards: the main board, which sits inside the glasses frame, and the sensor boards that connect through flex PCB cables

2017-imwut17-glabella-sensor_boards.jpg

Glabella's sensor boards incorporate a photodiode, a corresponding and spectrum-matching green LED, and a small opamp circuit. The PSoC-integrated ADC on the main board digitizes all signals in synchrony. (Left) Custom FPC cables connect the two sensor boards in the frame to the main board. (Right) Thin magnet wires connect the sensor board that is embedded in the nose pad to the main board, routed along the frame for minimal visibility.

2017-imwut17-glabella-sensor_board_underfill_protection.jpg

To limit the exposure of our sensor boards to the user's skin and protect them from external influences, such as sweat, we underfilled all components using fine traces of hot glue that we manually reflowed. Using an underfill keeps the main components of the sensor board exposed to make direct contact with the wearer's skin, such that the LED and the photodiode optimally inject light into the skin and observe optimal reflections, respectively.

2017-imwut17-glabella-sensor_pulse_bandpass_filtered.png

After applying a 0.4 Hz–8 Hz bandpass filter to the raw intensities, the pulse reflections exhibit clearly their characteristic features, which we consequently detect for extracting timestamps.

2017-imwut17-glabella-pulse_reflections_to_pulse_transit_time.png

Two pulse signals over the course of 16 seconds, from which we extract pulse transit times on a beat-by-beat basis. If the feature points around a pulse in the continuous stream of reflections fails our quality control, we skip the computation of the pulse transit time for this beat. In this example, the PTT curve shows the behavior of a person standing up wearing the prototype a er leaning back for 5 minutes. This produces a quick drop in relative systolic blood pressure, which quickly recovers to the previous values.

2017-imwut17-glabella-study_apparatus-task-in_the_wild.jpg

Each participant recorded values using two interfaces during the study. Our Glabella prototype glasses continuously recorded the user's pulse reflections at three sites on their head. Three times an hour, participants took a blood pressure measurement using a commercial cuff-based wrist-mounted radial blood pressure monitor. Each user participated in the study for a duration of five days.

2017-imwut17-glabella-study_results-correlation-pulse_transit_time-systolic_blood_pressure-angular_sta.png

The top figure shows the correlations of 1/PTT and systolic blood pressure between the angular artery and the superficial temporal artery, as well as the correlation coefficient for a linear model for each of the four participants.  The bottom figure shows the Bland-Altman diagram of the errors produced by the linear model in predicting systolic blood pressure values based on pulse transit times along the ground-truth values manually recorded by the monitor.