Marvin Wenzel, Nils C. Albrecht, Dominik Langer, Markus Heyder, Alexander Koelpin
©SHUTTERSTOCK.COM/RA2 STUDIO
Vital signs are most essential in medical care to evaluate a patient’s health situation. Today, electronic devices reliably measure the most important vital signs: heartbeat and respiration. Electrocardiography (ECG) measures the electrophysiological heart activity, while breathing can be detected with impedance pneumography (IP), which evaluates the periodic change of thorax impedance as we breathe in and out. While continuous monitoring of vital signs provides valuable medical information that can improve diagnostics, enable adaptive therapy, and generate new medical knowledge, there are also disadvantages. Medical procedures like ECG or IP require cables and electrodes, thus considerably reducing a patient’s mobility. Furthermore, electrodes cause skin irritation if applied too long, which often limits the continuous monitoring of vital signs to situations of absolute medical necessity.
An alternative constantly gaining attention is the contactless measurement of vital signs using radar. The first radar systems were built during the first half of the last century to detect movements of large objects, like ships or airplanes. However, advances in monolithic microwave integrated circuits (MMICs), 3D printing, and signal processing with deep neural networks allow one to design and build highly sensitive radar sensors today. Medical radar sensors sense microscopic motions caused by respiration and heartbeat to detect vital signs.
To encourage new ideas and introduce students to vital sign sensing radars, the IEEE Microwave Theory and Technology Society Biological Effects and Medical Applications Committee sponsored the 2022 Student Design Competition (SDC) for High-Sensitivity Motion Sensing Radar. This article gives an outline of the winning radar sensor presented at the 2022 SDC during the International Microwave Symposium (IMS) in Denver, CO, USA. We explain the competition scenario and figures of merit (FOMs), describe our system design process, and finally present the results achieved during the competition. Overall, our system demonstrated highly accurate vital sign detection of seated persons while balancing portability (e.g., size, weight, and power consumption) and cost.
The goal of the 2022 SDC was to design, fabricate, and demonstrate a radar system that can detect the cardiopulmonary activities of a seated person at rest who is breathing normally [1]. The general setup is shown in Figure 1. The subject is located 1 m from the radar sensor with another 1.5 m in the line-of-sight direction behind the subject, where no moving clutter is allowed. In contrast to previous competitions, e.g., those of [2], [3], and [4], the measurand is not a fixed physical quantity, like frequency or distance variation, that is known to the judges. Consequently, a reference sensing device is required to provide a ground truth for vital signs detected by the radar. Each team was required to prepare a reference sensing device synchronized with its radar sensor. The minimum requirement was to detect the fundamental frequencies of heartbeat and breathing. The scoring framework, as stated in the competition rules [1], is described in detail in the following sections. In total, a score of 100 is possible, with a maximum score of 20 per FOM. \[{\text{SC}}_{\text{total}} = {\text{SC}}_{\text{accuracy}} + {\text{SC}}_{\text{power}} + {\text{SC}}_{\text{weight}} + {\text{SC}}_{\text{potential}}{.} \tag{1} \]
Figure 1. The general measurement setup for vital parameter estimation as defined for the 2022 SDC. A person is seated 1 m distant from the radar sensor with an additional 1.5 m of clutter-free space behind the person. An electrode-based reference sensor provides a vital sign ground truth to analyze the radar performance. A computer for postprocessing connects to the reference sensor via USB, while the radar sensor streams its data wirelessly via Bluetooth (BL). A GUI displays the data in real time to select the optimal signal segment for final scoring.
${\text{SC}}_{\text{accuracy}}$ has a maximum of 40 points as it already includes the normalized and combined score for heartbeat and breathing accuracy (see the section “Vital Sign Accuracy” for more information). ${\text{SC}}_{\text{power}}$ rates power efficiency, while ${\text{SC}}_{\text{weight}}$ awards low-weight designs. ${\text{SC}}_{\text{potential}}$ assigns up to 20 points for the potential impact of the realized concept on a broader view.
DC power consumption was measured as the product of actual supply voltage and current. No power transfer from the connected laptop or batteries was allowed to ensure comparable power measurements. Four different scores were possible for power consumption and system weight, as shown in Table 1. By using integrated components, duty cycling, and a lightweight 3D-printed horn antenna, we achieved the highest possible score in both categories.
Table 1. The scores for the power consumption P DC and overall system weight W g. Our system achieved the maximum score of 20 in both categories.
To foster innovation, this year’s competition introduced a new Potential Score in addition to the FOM weight, power consumption, and accuracy, which were also used in previous competitions ([2], [3], [4]). This category allowed us to earn additional points for innovative ideas that might have been neglected in favor of system weight or power consumption. The maximum score for this category that the judges can award is 20.
The score for vital signs was based on averaged accuracies for heartbeat ${AA}_{H}$ and respiration ${AA}_{R}$. These were calculated as the error between the peak-to-peak intervals measured with radar and our reference sensor [(2) and (3)]. Since the competition rules lacked a normalization of the accuracies to the number of intervals measured, a higher heart rate could have resulted in a higher score. Therefore, a normalization factor $\left({20/I}\right)$ was added during the competition. \[{AA}_{R} = \frac{20}{I}{\cdot}\mathop{\sum}\limits_{{i} = {0}}\limits^{I}{\left({{1}{-}\frac{{|}{PPI}_{\text{R},\text{ref},\text{i}}{-}{PPI}_{\text{R},\text{radar},\text{i}}{|}}{{PPI}_{}}}\right)} \tag{2} \] \[{AA}_{H} = \frac{20}{I}{\cdot}\mathop{\sum}\limits_{{i} = {0}}\limits^{I}{\left({{1}{-}\frac{{|}{PPI}_{\text{H},\text{ref},\text{i}}{-}{PPI}_{\text{H},\text{radar},\text{i}}{|}}{{PPI}_{\text{H},\text{ref},\text{i}}}}\right)}{.} \tag{3} \]
With this normalization, a maximum score of 20 is possible for heartbeat and breathing detection accuracy, while the combined accuracy score ${\text{SC}}_{\text{accuracy}}$ is limited to 40. \[{\text{SC}}_{\text{accuracy}} = {AA}_{\text{R}} + {AA}_{\text{H}}{.} \tag{4} \]
Each team was provided 25 min to perform at least one measurement and choose a consecutive 60-s interval to calculate their accuracy score. The best segment could be chosen. However, respiration and heartbeat detection had to be based on the same interval.
With the presented scoring system, it is impossible to formulate a primary design goal as all FOMs are equally weighted. Consequently, our first design goal was to build hardware that achieves maximum weight and power consumption scores. Furthermore, we considerably reduced our bill of materials (BOM) for a real low-cost solution and implemented a Bluetooth interface to extend the portability of our sensor and earn additional potential points. Finally, to maximize the accuracy score, we decided to employ neural networks trained with publicly available data of radar-recorded vital signs [5].
In the following sections, we describe the design process and choices and explain the building blocks of our system in hardware and software. One of the most fundamental design choices for a radar system is the frequency of operation. We chose the 61-GHz Industry, Science, Medicine band for our radar system. The availability of a fully integrated radar front end in combination with reduced antenna sizes that scale down with wavelength allowed a considerable weight reduction of our design compared to a 24-GHz radar system.
Another design choice is the radar architecture. Most prevalent for vital sign monitoring are continuous wave (CW) and frequency-modulated CW (FMCW) radar sensors. Comparing the available integrated radar front ends, we found that FMCW sensors usually come with an increased power consumption. To achieve our goal of minimal power consumption, we chose a CW radar for our system.
A CW radar transmits electromagnetic waves at a specific frequency via the transmitter (Tx) antenna. The transmit signal ${s}_{\text{TX}}{(}{\text{t}}{)}$ propagates to the target and is reflected back toward the radar, where the received signal can be written as an attenuated and delayed version of ${s}_{\text{RX}}{(}{t}{)} = \kappa{\cdot}{s}_{\text{TX}}{(}{t}{-}\Delta{t}{)}$. A target’s motion changes the path length and, therefore, the delay $\Delta{t}$ between both signals. The delay corresponds to a phase shift $\Delta{\phi}{(}{t}{)}$, which is evaluated by the receiver (Rx). The relative distance change d (t) related to the detected phase variation $\Delta{\phi}{(}{t}{)}$ can be written as \[{d}{(}{t}{)} = {0}{.}{5}{\cdot}{\lambda}_{\text{RF}}{\cdot}\frac{\Delta{\phi}{(}{t}{)}}{{2}{\pi}}{.} \tag{5} \]
The relative distance change d (t) is unambiguous in an interval of $0.5{\cdot}{\lambda}_{\text{RF}}$ if no further processing, e.g., phase unwrapping, is applied.
Radar detects respiration by measuring the periodic distance change between radar and thorax as we breathe. Heartbeat detection, however, is more challenging to implement and interpret since a radar cannot measure the electrophysiological heart activity recorded by ECG. Instead, the mechanical heart activity is measured by detecting pulse waves or heart sounds. The pulse wave is generated as the heart contracts and ejects blood into the vascular system. Changes in pressure and blood volume propagate as pulse waves along the arteries [6]. Depending on the measurement position, volume changes of superficial blood vessels cause distance variations of up to several hundred µm that can be detected with a radar system [7]. Furthermore, it is possible to detect heart sounds with radar; these are vibrations related to the closure of heart valves during the cardiac cycle [8]. These vibrations propagate through the body toward the skin surface. The related distance change is only in the range of 10–15 µm, which is considerably less than for the pulse wave [9]. Nonetheless, their frequency range allows an efficient separation from respiration movement, which is why they were selected as additional vital sign and heartbeat indicators.
To provide a ground truth for heartbeat and respiration, we chose an electrode-based biosensor (MAX30001 by Maxim Integrated) that can record both biopotential (using ECG) and bioimpedance (using IP) simultaneously. Heartbeat and respiration signals are recorded synchronously. This allows us to synchronize the radar and reference sensor based on the first heartbeat detected by ECG and radar. The lab computer records data from the reference sensor for further processing. The respiration is found by performing a simple peak search within the bioimpedance signal. An implementation of the Pan–Tompkins algorithm is used to detect heartbeats from an electrocardiogram [10]. The detected locations of heartbeats and breaths serve as the ground truth to calculate the error function of the radar-based vital signs.
To achieve maximum scores for weight, power consumption, and potential, these aspects must already be considered in the system design process. To serve all categories, we pursue the goal of keeping the complexity and number of components as low as possible. For this reason, the presented system mainly consists of two components (see Figure 2): a radar MMIC and a microcontroller unit (MCU) with an integrated Bluetooth transceiver. The MMIC is a fully integrated 61-GHz radar front end from Infineon. It contains the RF circuitry: a voltage-controlled oscillator and a phase-locked loop with an external crystal oscillator. This allows it to generate a pure and stable CW tone. The generated signal is internally amplified by a power amplifier before being radiated by an integrated antenna in package (AiP). A 3D-printed horn antenna is mounted above the AiP to compensate for its large field of view. This increases the signal-to-noise ratio (SNR) and reduces sensitivity to clutter and interference. A detailed description of the horn antenna can be found in the section “3D-Printed Antenna.” The transmitted signal is reflected by the target and received by the horn antenna, which directly couples it to the chip’s integrated receive antenna. The received signal is down-converted using an IQ mixer and then low-pass filtered and amplified. A sample-and-hold circuitry keeps the baseband signals stable so that the chip’s RF components can be turned off between successive measurements. This way, a sampling rate of 2 kHz can be achieved with a low on-time of only 5 µs, keeping the power consumption to a minimum.
Figure 2. Block diagram of the proposed radar system: on the RF side a radar transceiver with an integrated AiP is used. A junction ensures optimal coupling between the AiP and waveguide feed for the 3D-printed horn. The complex baseband signals I and Q are sampled by the MCU and transferred via BL to a computer for further processing. PMICs provide optimized power supply for both the MMIC and MCU. AiP: antenna in package; BT: Bluetooth; PMIC: power management IC; ADC: analog-to-digital converter.
The baseband signals are converted to the digital domain using the MCU’s internal analog-to-digital-converter (ADC). A series of samples is combined into a Bluetooth packet, which is sent to the computer via the MCU’s integrated transceiver. In between transmissions the controller is put into sleep mode to reduce its power consumption. Disabling the RF components and using the MCU’s sleep mode reduces the mean power consumption to a minimum of 9 mW, as demonstrated by the power measurements in Figure 3.
Figure 3. With an average current of 3.17 mA and an average supply voltage of 3.07 V, an overall dc power consumption of 9 mW was reached. BT Con. Int.: BL Connection Interval.
Detecting microscopic movements related to vital signs requires a high SNR and is closely linked to a good antenna design. According to its radiation characteristics, the antenna projects a power distribution (spot) onto the subject’s upper body. The received signal represents an integration over the spot area, where each movement within the spot is weighted with the local power density. To combine maximum vital sign displacement and high local power densities for good sensitivity, a small spot focusing on the chest region, ideally close to the heart, is required. A spot radius of about 15 cm is necessary to project 50% of the radiated power on the subject’s chest. At a 1-m distance, this corresponds to a half-power beamwidth (HPBW) of less than 17°. Additionally, the sidelobe levels should be at least 10 dB below the main lobe for sufficient clutter rejection. To compensate for high free-space path losses of about 68 dB at 61 GHz, the antenna needs a gain of at least 20 dBi.
The radar front end features integrated patch antennas (Rx and Tx) with only 6 dBi gain. At 61 GHz, aperture radiators with the required characteristics have compact dimensions and low losses. Therefore, we designed horn antennas for the Tx and Rx paths to retain the bistatic design of the integrated antennas. Without the opportunity to bypass the AiP, a transition to a rectangular waveguide (RWG) was developed to feed the horn antenna directly via the AiP. The transition consists of a specially shaped RWG positioned above the AiP with a mounting structure. It ensures optimal coupling into/from the RWG, which connects the AiP and Tx/Rx horn antennas.
A pyramidal horn was developed in CST Microwave Studio. The aperture of 24 mm × 17 mm generates a simulated gain of 21 dBi with an HPBW of 15.2° in the H-field plane and 14.8° in the E-field plane. The sidelobe suppression is at least 11 dB. Horn antennas are usually made of aluminum or brass, which are incompatible with our weight limit. However, a low-loss propagation of the wave can also be achieved with a 1.5-µm copper layer corresponding to five skin depths at 61 GHz. The copper can be applied to a lightweight dielectric carrier to increase the mechanical stability. For easy manufacturing, a 3D CAD model of our horn antenna was printed using a Form3+, a stereolithographic 3D printer from Formlabs. After printing, we applied a thin silver layer to the dielectric resin model using Tollens’ reagent. This is necessary as our copper-plating process requires a conductive surface. Next, the antenna was electroplated with a 1.5-µm copper layer. An overview of the fabricated system is shown in Figure 4.
Figure 4. An overview of the fabricated system with the 3D printed antenna, backside of the PCB and its components, and a frontside schematic indicating the RF MMIC and antenna mounting. Without the horn antenna, the overall system is only 22 by 44 mm and weighs 1.7g (with horn antenna 7.8g).
The proposed 3D-printed and copper-plated bistatic antenna weighs only 5.9 g. A comparable antenna made of brass weighs approximately 26 g—this corresponds to a weight reduction of 77%. Finally, the antenna pattern was measured in an anechoic chamber to confirm the simulated performance metrics. The results are shown in Figure 5 and have excellent agreement between simulated and measured radiation characteristics.
Figure 5. Normalized farfield pattern of the manufactured 3D-printed horn antenna: simulated () and measured () E-plane and simulated () and measured () H-plane pattern.
As described in the section “System Description,” our baseband signals are digitized by the MCU and directly transmitted to the lab computer via the integrated Bluetooth interface. At the same time, reference sensor data are streamed to the computer via USB to provide a ground truth for our scoring algorithm. After the measurement, both data streams are processed in a Python-based GUI that also visualizes the signals. Our proposed processing contains multiple steps:
Since radar and reference sensor measurements are started simultaneously, we implemented a simple but effective synchronization based on the temporal alignment of the first detected heartbeat in radar and ECG data.
Next, we filtered the reference as well as the radar data. The ECG and IP data were filtered with a low-pass and high-pass filter to remove power line interference and baseline drift, respectively. The radar data was processed with fourth-order Butterworth bandpass filters of different frequency ranges to separate the respiration (0.05–0.7 Hz) from pulse wave and heart sound components located at higher frequency bands (2–80 Hz). Next, we extracted the vital signs from the reference and radar data. The algorithmic approaches we chose for respiration and heartbeat detection are different and are therefore described separately.
We use the IP reference signal and a low-pass filtered radar signal for respiration detection. With physiological breathing rates of 12–20 breaths/min (maximum frequency 0.33 Hz), signal components above 0.7 Hz are considered unrelated to respiration. We apply arc tangent demodulation to the complex radar baseband signal to calculate the distance signal from the phase variation. Finally, respiration is detected by performing a peak search on the demodulated distance signal. For our ground truth, we apply a peak search algorithm directly to the IP data to find the maxima of the thorax impedance, which are correlated to inspiration. The radar-based respiration signal is shown in Figure 6(a) and shows excellent agreement with the IP-based respiration signal.
Figure 6. (a) Measured respiration signal by the reference bioimpedance IP () and by the radar system (). (b) Measured ECG signal () and the prediction output of the neuronal network ().
Heartbeat detection is more complex as the displacements of pulse waves and heart sounds (in micrometers) are several orders of magnitude smaller than those of respiration (in millimeters). Therefore, we use a neural network to predict a normalized heartbeat probability from our radar data. The prediction is based on signal features extracted from bandpass-filtered radar data. Our set of features has previously been shown to be related to the heartbeat and includes a Hilbert envelope and the demodulated distance signal [11], [12]. To increase the training and classification speed, we perform down-sampling by a factor of 20 (signals originally sampled at 2 kHz) and normalize our features by calculating their z-score. For the prediction, we use a long short-term memory (LSTM) architecture, which allows us to include information from previous signal segments. Furthermore, a bidirectional layer is used to also process radar data in the reversed temporal direction, thus including future information. Adding this context information is especially useful to process time series and has already been successfully applied for vital sign detection in radar data [11]. The neural network also includes a dropout layer to reduce overfitting risk and a dense layer with a nonlinear activation function. The neural network was trained using radar-recorded vital signs of multiple persons to obtain a generalized model and avoid overfitting to a specific person.
The ground truth for predicted heartbeats is found by applying a Pan–Tompkins algorithm to the preprocessed ECG data [10]. The predicted heartbeats and ECG ground truth show excellent alignment across all detected beats, as shown in Figure 6(b). Furthermore, Figure 7 shows the instantaneous heart rates calculated from ECG and radar for a 60-s interval. The radar-based instantaneous heart rate is shown here as a raw signal, and some outliers can be observed. These outliers originate from the calculation that is updated for each predicted heartbeat. If a predicted heartbeat occurs too early, the shorter beat-to-beat interval increases the instantaneous heart rate. Similarly, the following beat-to-beat interval will be longer, thus decreasing the instantaneous heart rate. These rapid changes of the heart rate lead to observable outliers that could easily be removed, for example with a smoothing filter. Nonetheless, both heart rates show good agreement across the entire segment.
Figure 7. Instantaneous heart rate measured by the reference ECG () and by the proposed radar system (). BPM: beats per minute.
Finally, the accuracy score for heartbeat and respiration is calculated, as described in the section “Vital Sign Accuracy.” The score is maximized by finding the optimal consecutive 60-s interval within our measurement as proposed in the competition rules [1].
We presented a lightweight, low-power 61-GHz CW radar system that accurately detects heartbeat and respiration and can be built for less than 15 € (see Table 2). Our design was specifically tailored to the requirements of the 2022 SDC for high-sensitivity motion sensing radar to maximize the score. Integrated components and a 3D-printed antenna with copper plating allowed us to build a radar system with an overall weight below 10 g. Nonetheless, our antenna achieves a high gain and narrow beam with HPBWs in the E- and H-planes of 14.8° and 15.2°, respectively. We used extensive duty cycling for our MCU and utilized the on-demand activation of RF blocks in our front end to reach an overall dc power consumption of 9 mW. To demonstrate the portability of our system, we implemented a Bluetooth interface that streams the radar data to a laptop for further processing. Theoretically, this feature allows a battery-powered, independent operation if the processing algorithms are run on a digital signal processor. During the competition, we measured and automatically detected heartbeat and respiration using our CW radar system and a pretrained neural network with an LSTM architecture. Both vital signs showed excellent agreement with the reference ground truth provided by ECG and IP. Thus, we demonstrated that low-cost, lightweight, and low-power radar systems can be used for reliable, contactless beat-to-beat heart and respiration monitoring.
Table 2. The BOM for a single sensor system. With careful design choices, the overall BOM was reduced to <15€.
This work was partly supported by the German Research Foundation (DFG, German Research Foundation) under Grant SFB 1483, Project ID 442419336.
[1] F.-K. Wang, J.-M. Munoz-Ferreras, C.-T. M. Wu, and O. Boric-Lubecke, “IMS-2022 student design competition instructions: High-sensitivity motion sensing radar,” Institute of Electrical and Electronics Engineers, Piscataway, NJ, USA, 2022. Accessed: Aug. 18, 2022. [Online] . Available: https://ims-ieee.org/sites/ims2019/files/content_images/SDC2022v3.pdf
[2] F. Lurz, F. Michler, B. Scheiner, R. Weigel, and A. Koelpin, “Microw(h)att?! Ultralow-power six-port radar: Realizing highly integrated portable radar systems with good motion sensitivity at relatively low cost,” IEEE Microw. Mag., vol. 19, no. 1, pp. 91–98, Jan./Feb. 2018, doi: 10.1109/MMM.2017.2759640.
[3] F. Michler, B. Scheiner, F. Lurz, R. Weigel, and A. Koelpin, “(Micro)metering with microwaves: A low-cost, low-power, high-precision radar system,” IEEE Microw. Mag., vol. 20, no. 1, pp. 91–97, Jan. 2019, doi: 10.1109/MMM.2018.2875612.
[4] B. Scheiner, F. Michler, F. Lurz, R. Weigel, and A. Koelpin, “Nothing beats SNR: Single-digit micrometer ranging using a low-power CW radar featuring a low-weight 3D-printed horn antenna,” IEEE Microw. Mag., vol. 21, no. 1, pp. 88–95, Jan. 2020, doi: 10.1109/MMM.2019.2945150.
[5] S. Schellenberger et al., “A dataset of clinically recorded radar vital signs with synchronised reference sensor signals,” Scientific Data, vol. 7, no. 1, Sep. 2020, Art. no. 291, doi: 10.1038/s41597-020-00629-5.
[6] J. Alastruey, K. H. Parker, and S. J. Sherwin, “Arterial pulse wave haemodynamics,” in Proc. 11th Int. Conf. Pressure Surges, 2012, pp. 401–443.
[7] C. Will et al., “Local pulse wave detection using continuous wave radar systems,” IEEE J. Electromagn., RF, Microw. Med. Biol., vol. 1, no. 2, pp. 81–89, Dec. 2017, doi: 10.1109/JERM.2017.2766567.
[8] R. F. Schmidt and F. Lang, Ed., Physiologie Des Menschen. New York, NY, USA: Springer-Verlag, Oct. 2007.
[9] C. Will et al., “Radar-based heart sound detection,” Scientific Rep., vol. 8, no. 1, pp. 11551, Jul. 2018, doi: 10.1038/s41598-018-29984-5.
[10] J. Pan and W. J. Tompkins, “A real-time QRS detection algorithm,” IEEE Trans. Biomed. Eng., vol. BME-32, no. 3, pp. 230–236, Mar. 1985, doi: 10.1109/TBME.1985.325532.
[11] K. Shi et al., “Contactless analysis of heart rate variability during cold pressor test using radar interferometry and bidirectional LSTM networks,” Scientific Rep., vol. 11, no. 1, 2021, Art. no. 3025, doi: 10.1038/s41598-021-81101-1.
[12] D. B. Springer, L. Tarassenko, and G. D. Clifford, “Logistic regression-HSMM-based heart sound segmentation,” IEEE Trans. Biomed. Eng., vol. 63, no. 4, pp. 822–832, Apr. 2016, doi: 10.1109/TBME.2015.2475278.
Digital Object Identifier 10.1109/MMM.2022.3226546