A portable, low-power acoustic device for accurate invasive blood pressure monitoring (2023)

idea and prototype

Early methods of aligning pressure transducers to the proper height included qualitative inclinometers fixed between the patient and the transducer or mechanical devices with wheels that moved from the patient's hydrostatic reference point to the pressure transducer.^11,12.Optical techniques have been described to direct light from the height of the pressure sensor to the patient's hydrostatic reference point or retractable inclinometer for optimal positioning^13,14,15,16However, it requires the provider to manually readjust the transducer when the patient's reference point changes. An optical distance measuring device placed under the bed can be used to measure the height of the bed from the floor. However, the head of the bed can be raised or lowered and the patient can move around in the bed, causing changes in the height of the hydrostatic reference point that cannot be estimated by such devices. Finally, one can imagine a camera-based system that can provide detailed information, but camera systems raise significant privacy issues in a clinical setting.

Our project has two main hardware components: a low-power wireless wearable and a tracking device. The wearable device is a 2.5 cm diameter patch that is attached to the patient and can be used to determine their spatial position from a tracking device connected to an IV pole at a fixed height (Fig. 1).1one two). The tracking unit does not interfere with the clinical use of the IV brace as it fits over the hook used to hang the IV fluid bag using the 3D printed loop. The locator consists of a speaker array with four speakers and a Bluetooth receiver (Figure 1).2ONE). The locator emits an inaudible frequency modulated continuous wave (FMCW) signal that is picked up by a microphone on the handheld device (Figure 1).2b).

To determine the elevation of the hydrostatic reference point, the distance between the handheld microphone and three of the four location speakers was calculated. Triangulation is then used to calculate the 3D position of the mobile device by extracting time-of-flight information from the FMCW audio signal transmitted by the locator. Calculating flight time accurately is difficult for two main reasons: (1) Calculating flight time requires two devices - a locator and a handheld - to be continuously synchronized with each other to share a common clock. Given the precision required for positioning, this is not trivial to achieve, as it only takes 0.03ms for the sound to travel 1 cm. (2) Acoustic signals transmitted by the receiver bounce off nearby surfaces, objects and people before reaching the microphone. This is known as multipath and makes it difficult to separate the line-of-sight path from all reflected paths.

The above challenges are addressed using various technology elements, as shown in Figure 1.2to do. First, a combination of Bluetooth-based synchronization (BLE) algorithms and post-processing is used for precise time synchronization between the two devices. Specifically, the tracker transmits synchronized Bluetooth packets at a rate of 200 Hz, which contain information about its free running clock. When the mobile device receives this packet, it will synchronize its own clock. While this Bluetooth technology can significantly reduce clock drift over time, it does not eliminate it due to sampling and accuracy limitations. A post-processing algorithm solves this problem by performing linear regression on elevation estimates over longer periods to calculate residual displacement after Bluetooth synchronization. The received acoustic signal is then resampled using software for time alignment with the signal transmitted to the receiver (see Methods for details).

To solve the multipath problem and obtain time-of-flight calculations on a subcentimeter scale, we rely on recent work in¹⁷A phase-based FMCW algorithm is introduced to account for direct path versus multipath reflections. However, this plan¹⁷Focus on monitoring a single speaker in a microphone array. Instead, the current design has the opposite configuration: the handheld has a low-power microphone to reduce power consumption, while the plug-in finder uses a more sophisticated speaker array and is able to calculate readings. of altitude. The algorithm is adapted to this inverse problem to obtain an accurate height calculation for mobile devices. Finally, several techniques are used to minimize energy consumption and deal with the movement of people in the environment.

Mobile device hardware design

Our mobile device includes a pulse density modulation (PDM) microphone (Invensense ICS-41350), a Bluetooth Low Energy (BLE) microcontroller (Nordic nRF52840), and an ultra-low power accelerometer (Bosch BMA400). The device is powered by a CR2032 coin cell battery. The device's built-in PDM microphone is set to a clock frequency of 3.2 MHz. The internal decimation rate of the PDM is 64, which gives a sampling frequency of 50 kHz. As the adult human ear (18–22 kHz) cannot detect nearby ultrasonic frequencies, a sampling rate of 50 kHz is high enough to avoid distortion.

Two 16-bit pulse code modulation (PCM) buffers of sample size 512 are used cyclically: one is filled with incoming PCM data, while the other is processed. The DMA is responsible for timing and converting PDM data to PCM. One buffer is always connected to DMA while the other buffer is released to handle the rest of the data line.¹⁸.When the buffer attached to DMA is full, the buffers switch roles and we start processing the data in the buffer that was just released and append another buffer back to DMA. With this design, we can always operate with a continuous PCM stream.

To maximize your device's battery life, we've also integrated an ultra-low power accelerometer. During periods of device inactivity, we use the accelerometer to turn off the device's audio and Bluetooth subsystems. In this idle state, we configured the BMA400 to operate in low power mode, consuming only 160 nA. In this mode, we configure the BMA400 to stop motion along the z-axis so the wearable can activate Bluetooth and microphones while the bed or patient moves. This way we alternate between a fully enabled state (Bluetooth, accelerometer, microphone on) and a low power state (accelerometer only). The power consumption of mobile device components is shown in the table1.The estimated battery life is 36 hours (6mA) and 534 days (17.53µA) in full power and low power mode, respectively.

A wireless handheld device transmits PCM microphone and accelerometer data to a tracking device to calculate altitude information. However, to maximize performance, we use the highest Bluetooth rate and packet size supported by Bluetooth 5.0, which are 2 Mbps and 247 bytes respectively.

The hardware schematic and mobile device layout were designed using the KiCad open source eCAD tool. PCBWay manufactures and assembles a 2-layer flexible printed circuit board. The 3D printed shell was designed using AutoDesk Fusion 360 and printed using a Phrozen Sonic Mini using a liquid resin manufacturing process. The MEMS microphone is behind an exposed cover on the outer surface of the wearable. Finally, we designed a universal mounting port on the opposite side of the wearable so that our system can be connected to commonly used ECG electrodes in clinical settings.

Locating the hardware project

The positioner transmits FMCW chirps from its 4 speakers, which are spaced geometrically at the corners of the square. In addition, the tracker is responsible for establishing a wireless connection with the wearable patch to receive data from the microphone closest to the patient's chest. Four FMCW chips are generated using a Teensy 3.6 microcontroller and physically sent to a Cirrus Logic CS42448 encoder over a digital time division multiplexed (TDM) audio bus. Although I2S is a standard stereo digital audio interface, our system requires 4 channels, so we choose TDM which supports up to 8 channels. Once the audio signal hits the encoder, it outputs single-ended analog audio to each of our speakers. We used the nRF52840 for time synchronization between the wearable and the tracker and a Raspberry Pi 4 for calculations.

The locator is synchronized with the wearable using a time sync beacon provided by the Nordic Library¹⁹.Microphone data transmission and time synchronization work independently. Time synchronization works by transmitting the finder's timestamp to the mobile device, which can set its own clock to synchronize with the finder's clock. Using synchronized clocks, we know the exact time of each recorded audio sample, which is critical for accurate position estimation.

Finally, the entire system was mounted on a 3D-printed template compatible with standard IV poles, as shown in the accompanying Figure 1.1, so that our locator can receive reliable signals from the wearable. The top part of our puzzle (Fig.1)Let the positioner rest and close the attachment for stability. The cuff and positioner can then be moved to any standard IV pole, the bottom of our cuff matches the top of the IV pole. While the clamp is attached to the IV pole, we have also included Velcro as an added measure to keep the retainer attached to the IV pole.

Time synchronization algorithm

A key requirement of our system is that the two devices - the tracker and the handheld - must be constantly synchronized with each other in order to share a common clock. Given the precision required for positioning, this is not trivial to achieve, as it only takes 0.03ms for the sound to travel 1 cm. We use a combination of Bluetooth-based synchronization algorithms and post-processing to achieve accurate time synchronization between the two devices.

Bluetooth-based time synchronization

In our system, both the handheld and the finder have a 32 MHz clock source with a frequency tolerance of ±20 ppm. So, in the worst case, both devices move 2.4 milliseconds per minute. To stay in sync, we use Nordic's TimeSlot API¹⁹, which gives us access to the underlying radio hardware between Bluetooth transmissions and gives us transmissions that transmit and receive beacons with precise time synchronization¹⁸Both the finder and our handheld maintain a free-running 16 MHz hardware timer with a maximum value of 800,000, which overflows and wraps at a rate of around 20 Hz. The tracker is set to the timer and the wearable synchronizes the free-running timer with the timer on the controller. The positioner (time controller) transmits time synchronization packets at a rate of approximately 200 Hz. These packets contain the idle timer value at the time of transmission of the radio packet. When the mobile device receives this packet, it synchronizes its own clock by adding or subtracting the offset from its own free-running timer. After establishing a common clock between the receiver and the handheld, the received audio signal is resampled to align with the time the receiver transmitted the signal and correct for the offset. Specifically, we iterate over the signal by performing an FFT and an inverse FFT using a base frequency shift:

hereONEis the displacement factor (ONE=⟩1 means no displacement),x_Atumis the original signal in the time domain and\({\hat{x}}_{t}\)is the resampled signal.

post-processing algorithm

Although the Bluetooth time synchronization technique described above can significantly reduce time drift over time, it does not eliminate time drift due to sampling and accuracy limitations. To further improve the accuracy, we designed a post-processing algorithm, which is applied to the signal after the Bluetooth time synchronization algorithm. We exploited the observation that, in a targeted clinical setting, body movement is sparsely distributed in the temporal domain over short periods of time. For example, during several hours of surgery, the patient moves only 10-20 times, each movement lasting a few seconds. Also, every 5 minutes much less than half of the duration is spent moving the body.

To take advantage of this, we keep a history of the last 5 minutes of height estimation data. We split the 5-minute data into 30-second blocks and use each 30-second block to compute the residual sum of squares after linear regression. We sort these residual values, select their lower half, and calculate the mean slope of the corresponding linear regression results. Average slope is an estimated drift factor, which is then used to compensate for future drift.

Algorithm for Altitude Calculation

A key component of automatic elevation change tracking is an acoustic 3D positioning system, in which the relative position of the handheld device and a locator attached to a pole are calculated in real time. A microphone on the wireless unit receives audio signals from multiple speakers on the tracker. We only use three of the four speakers for the signal and if one of the three isn't working properly, the others act as redundancy. Given the geometry of a speaker array, any three subsets of it can be used to obtain the 3D position, more specifically, the height difference between two devices.

ESPECIALLY,EUThe speaker emits a time-shifted version of the FM chirp signal:

Wherex_EU(Atum) is the speaker signalEUno timeAtumIn the time domain,Atumis the duration of the chirp, andcomer₀ecomer₁are the starting frequency and the ending frequency, respectively. Speaker hum is delayed multiple times\(\frac{T}{5}\).In our application we putcomer₀,comer₁eAtum18 kHz, 22 kHz and 43 ms, respectively.

The signal received by the wearable can now be written as the sum of all the speaker signals:

herepiONEAtumHSecond_EUis the set of all speaker propagation pathsEUOn the tracking device connected to the microphone on the mobile device.NoEUSecondAtum(EU,pi) is the distance from the speakerEUpath to microphonepi,ONE_piis the attenuation coefficient of the pathpi,to dois the speed of sound andno(Atum) is random noise, including material noise and environmental noise.

The wireless device sends this signal,Sim(Atum), using Bluetooth to return to the tracking device. The tracker uses a two-step process to estimate the relative change in distance from each speaker to the direct path to the microphone. The speaker layout is fixed, with each speaker located at one corner of the rectangular panel, as shown in Figure 1.2TO DO.

During setup, we perform an initialization step where we measure the initial position of the wireless device relative to the speaker array by placing the wireless device next to the speaker array at a specific known location (locator). We then compute the initial distance reported by the algorithm between each speaker and the wearable, denoted asMan_EU, ForEUfirst speaker. In its operation, the algorithm estimates the distance between the track and each of the three loudspeakers in operation as a function of time, expressed as\({\widetilde{d}}_{i}(t)\). Then get three corresponding absolute distances after time synchronization and post processing,\({d}_{i}(t)={{{{{{{\bf{post-processing}}}}}}}({{{{{{{{\bf{timesync}}}} }}} } }}({\widetilde{d}}_{i}(t)-{D}_{i}))\)and its average,\(\bar{d}(t)=\frac{{\sum }_{i}{d}_{i}(t)}{3}\).

The difference in distance measurements can then be used to estimate the 3D position of the wearable relative to the speaker array as follows:

Wheregrandeis the separation distance between pairs of adjacent loudspeakers andG(Atum) is an estimate of height.

Therefore, if we can accurately calculate the one-dimensional distance,\({\widetilde{d}}_{i}\)Over time, between the microphone and the three speakers, we can measure the 3D position of the wireless device. In the remainder of this section, we describe how to calculate the one-dimensional distance between the microphone and the speaker. The main challenge to accurately calculate the 1D position is that, in addition to the direct path from the speaker to the microphone, the microphone receives reflections from all objects and people in the environment (long-range multipath) as well as reflections from the patient. . body next to the wireless path (body multipath ). Therefore, our algorithm must accurately determine the direct path in the presence of all these reflections. To do this, we use a two-step process.

• Step 1: Use a high pass filter to remove distant reflections.Use the received signalSim(Atum) in each cycle of durationAtum, we first disconnect the four separate chirp signals coming from each speaker. This can be done by extracting the four peaks of the demodulated signal into the frequency domain using a discrete Fourier transform similar to the previous FMCW processing algorithm.¹⁷.for every chirpx_EU(Atum) by the speakerEU, we then apply a finite impulse response (FIR) filter in order to leave only a narrow band of frequencies around the peak. To do this, we adaptively vary the delay of the FIR filter using the signal-to-noise ratio (SNR) of the received audio signal, when the SNR exceeds 10 dB we set the delay to 15 ms, while at lower SNRs we use a longer delay. tax 30 ms. These parameters are not defined for a specific environment and adjusting them can further improve accuracy. Our assessments are made in real chambers. However, other locations with acoustic reverb devices or loudly interfering background sounds may lead to worse results.

• Step 2: Use the FMCW phase to remove bulk multipathThe above procedure separates the chirp signal received from each speaker and reduces the long-distance multipath effects of objects and people in the room whose distances are very different from the direct path. This leaves us with a residual indirect path of body reflections around the wireless device. The sum of the remaining indirect paths is less than the direct paths when there is no corresponding reflective obstruction (with the exception of clothing) between the wireless device and the receiver. Therefore, we can derive the distance corresponding to the forward path using the FMCW phase in the output signal from step 1. Specifically, the phase of the FMCW signal from the receiver can be written as¹⁷,

  $${{{{{\boldsymbol{\phi }}}}}}}}(t)\approx -2\pi \left(\frac{B}{T}t{t}_{d} + {f}_{0}{t}_{d}-\frac{B}{2T}{t}_{d}^{2}\direita)$$

  (8)

  hereAtum_Nois the arrival time of the direct route, andSecondis the bandwidth of the chirp signal, 4000 Hz. Therefore, for each speakerEU, by sampling the phase of the filtered signal corresponding to the loudspeakerEUThey result from the last step at a given timeAtum(For example.,Atum=Atum/2), we solve the quadratic equation above to derive possible solutionsAtum_No,EU.We calculatedAtum_No,EUFor direct paths like solutions in the FMCW chirping range, [0,AtumWe then compute the one-dimensional distance (with offset) between the speaker and microphone as\({\tilde{d}}_{i}=c{t}_{d,i}\), whereto dois the speed of the sound signal in air.

clinical trials

Written informed consent was obtained from 26 patients for this study, which was approved by the Institutional Review Board at the University of Washington (STUDY00009667). English-speaking patients undergoing cardiac surgery at the University of Washington who required invasive blood pressure monitoring were eligible and consented prior to surgery. Each patient was studied up to two times, once with an endotracheal tube and mechanical ventilation and once without an endotracheal tube with automatic ventilation. For research purposes, we use the following two converters:

clinical transducer

The transducer is attached to the patient's skin at the level of the heart (the hydrostatic reference point) and the tube is long enough to move up and down with the patient as they move around the bed. Although connecting the transducer directly to the patient is not recommended, in our clinical research we used it to provide real data.

fixed converter

This transducer attaches to a bedside IV pole, which is the desired configuration in these scenarios. It is initially placed at the lowest height of the bed and does not move when the patient's bed is moved.

We noticed that even when we placed two sensors at the same height, they didn't give the same reading. For this reason, fixed transducer measurements have been normalized to clinical transducers to explain the major differences in transducer measurements as follows:

herericeONEpi_{to do,0}is the average measurement with the clinical transducer attached to the patient's skin at the lowest height of the bed andriceONEpi_pi,0is the average reading of a fixed transducer.

Our algorithm adjusted the blood pressure measured by the fixed sensor to account for the additional hydrostatic pressure due to the height difference between the patient's hydrostatic reference point and the sensor. This can be done by multiplying the height difference (in centimeters) by a factor of 0.735 to convert to mmHg²⁰.

Specifically, we recorded three different mean arterial measurements as part of our clinical study.

clinical mean blood pressure

Clinical mean arterial pressure is the mean arterial measurement,riceONEpi_{to do}, which was used as part of clinical care during the study and came from a transducer attached to the patient to move as they move up and down in bed.

Mean arterial pressure laser measuring tool

The mean arterial pressure of the laser range tool is the calculated mean arterial pressure,riceONEpi_Elevator, using a fixed pressure sensor and a laser ranging tool. The change in distance is calculated from the change in the distance output of a laser tool located under the patient's bed. ifH_Elevator,xIndicates the height measured by the laser in cm,H_Elevator,0Indicates the elevation measured by the laser rangefinder at the lowest vertical setting in the ICU bed andriceONEpi_Secondis the pressure reading from the fixed sensor, then,

Automatic Height Tracker Mean Blood Pressure

Auto height tracker mean arterial pressure is calculated mean arterial pressure,riceONEpi_r.In this case, the distance between the static pressure reference points was calculated from vertical measurements taken from our own altitude monitoring system, and the wearable was placed on the patient's skin at a heightH_ONE,xof the tracking device.H_ONE,0Indicates the height of the wearable in its lowest vertical position in an ICU bed. Our estimated blood pressure is calculated as follows:

Statistics and Reproducibility

This study included 26 human participants. The mean age of patients was 56.8 years (range, 24 to 80 years), with a male to female ratio of 0.55. All patients underwent open-heart surgery and therefore had severe heart disease. As participants were in the intensive care unit, measurements were generally not repeated to minimize impact on patient care. A second set of data was obtained whenever possible (3 cases) if the patient was in the ICU the following day and the arterial catheter remained in place. Evaluate data using Excel and Python.

report summary

For more information on the study design, visitOrganic Portfolio Report SummaryLink to this article.