Remote Health Monitoring System for Bedbound Patients

In this paper, we present a novel solution for the remote breathing and sleep position monitoring by using a multi-input-multi-output (MIMO) radar. Our proposed system is able to monitor a number of people simultaneously, and therein we use a high-resolution direction of arrival (DOA) detection for finding closely separated targets. So, it effectively increases the number of target detection and reduces the cost by reducing the number of sensors. Furthermore, our proposed system is capable of identifying the sleep position of each monitored person by selecting appropriate target features and using a support vector machine (SVM) classifier. The breathing analysis involves designing an optimum filter for estimating both the breathing rate and the noiseless breathing waveform. In addition, we use the radar in a bedroom environment above a bed where two subjects sleep next to each other. The accuracy of the breathing monitoring subsystem is more than 97% for human subjects in the bedroom compared with a reference sensor. Also, the correct rate for sleep position detection is more than 83%.


I. INTRODUCTION
The COVID-19 outbreak has tremendously changed the world in many ways, but it teaches us many lessons to adapt and maintain sustainability in pandemics. The protection of healthcare providers must be the first priority. However, the exponential increment in the number of patients, in addition to the shortage of personal protection equipment (PPE) creates a huge challenge to control the pandemic. Body temperature, blood oxygen density, heartbeats, and respiration are the primary vital signs that are monitored constantly in hospitals for COVID-19 patients, but the conventional medical devices require attachments to the patient's body. Therefore, inevitably, healthcare workers should touch surfaces that COVID-19 patients have touched, increasing the risk of infection. So, we need to employ new technologies for long-distance monitoring of multiple patients. This reduces the exposure time and cost, and the large number of nursing staff needed to administer customized testing.
Sleep apnea, bedsores, sleep disorders, and in general, bedbound patients require various wearable or contact medical tools to record vital signs. Sleep monitoring with the polysomnography (PSG) as a gold standard is very costly and requires the continuous supervision of a physician. Mostly, these are contact devices intervening with the patients' freedom. In some cases, the devices are impossible to be connected to the patient; for instance -a patient with burned skin. Moreover, health monitoring is interrupted when the attachments become loose, or at least, the measurement error increases. However, more convenience and reliability can be offered by contactless devices. For instance, it has been shown that radars can detect vital signs without any contact with the body by capturing tiny chest motions due to cardiorespiratory activity [1]- [4]. Although they can facilitate non-obstructive, touchless, and long-term vital signs detection, practically using radars for heart rate and breathing rate (BR) detection lacks enough accuracy in the presence of body motions [1] or the radar platform motion [4]. In particular, they can be used for bedbound patients where they are supposed to be stationary most of the time with occasional movements.
For respiration rate estimation, the aim is to find the fundamental frequency of the periodic respiration signal mixed with noise and interference, so they are pseudo-periodic signals. This problem has been investigated extensively in literature under the topic of f 0 estimation, fundamental frequency estimation, and pitch detection. Extracting periodic signals from a single channel noisy observation is similar to blind source separation (BSS) problems; however, they differ in nature for having a periodic assumption on the behaviour of the sources. Furthermore, there are many parametric high resolution methods such as linear prediction [5], subspace methods [6], [7], harmonic fitting [8], maximum likelihood [9], and cepstral method [10]. Among subspace methods, MUltiple SIgnal Classification (MUSIC) analysis exploits the signal subspace models to enhance, separate, and estimate the periodic signal properties [11], [12]. Christensen et al.,in [13], did not only find the fundamental frequency, they also fully characterized the periodic signals. Here, we apply the high resolution optimum filter in [11] to estimate breathing rate and its noise-less waveform by receiving the radar complex signal. Also, the complex signal detection is more robust than the phase domain analysis since it does not need nonlinear phase computation, phase unwrapping, and also the vital DC cancellation does not create distortions.
In this work, we use a commercial millimeter-wave FMCW radar by Texas Instruments, AWR1243 [14], operating at the centre frequency of 79 GHz with 4 GHz sweeping bandwidth with high sensitivity to the body motions [15]. The radar has 3 transmitters and 4 receivers. For sleep position detection, we collect features from each detected target points as the inputs to support vector machine (SVM) classifier. The detection is novel and does not have any similar prior arts. Furthermore, we use a smart garment, Hexoksin vest, by Carre Technologies inc. as a reference sensor [16]. It can monitor BR, tidal volume (vt), minute ventilation, and hip motion intensity (HMI) among other vital signs with higher accuracy than 98% [17].
In the next section, we will discuss the proposed system architecture. In Section III, we will give the measurement results of experiments. In Section IV, the paper is concluded with remarks and highlights.

A. MIMO Radar
Our proposed radar processing involves many steps with the major processing blocks sketched in Fig. 1. The MIMO radar is based on a set of orthogonal transmit waveforms [18]. A particular orthogonal transmission scheme is time division multiplexing (TDM), which creates orthogonality in time by sending signals from only one transmitter at a time. Therefore, at the receiver, the transmitted signals from each transmitter can be separated by knowing that the Rx recording in each time interval corresponds to the transmission from a particular Tx antenna. This separation helps to copy the receiver array based on the transmit element locations. For instance, in Fig.  1b, the transmit array has 2 elements along azimuth and the receiver has 4 elements. So, the measurement in the second transmission has an additional phase shift with respect to the first chirp depending on the distance of the two transmit antennas. If the distance of two transmitters is equal to the width of the receiver array, so the measurement from the second transmission at the receivers is equivalent to placing virtual Rx elements next to the actual Rx array -the blue array virtually extends the receiver due to the transmission from the second channel. Theoretically, the virtual 8-element receiver array has the angular resolution of about 2/N = 14.23 degrees, meaning that statistically the same targets are resolvable at 3 meters only if they are apart at least 70 centimeters.

B. Target Detection and Breathing Analysis
Every other chirp is transmitted from a Tx antenna, so the extended array is formed by treating the measured samples of the odd chirps as received by the actual receivers, and the sample of even chirps as received by the virtual receivers. So, at each particular sample index, we have 8 samples across all channels -the actual and virtual channels. The received samples within a chirp period, are then stacked vertically, and the samples across multiple chirps, i.e. slow-time samples, are stacked horizontally. This makes a 3D array called radar cube as shown in Fig. 1b. An FFT is applied to the chirp samples, the vertical direction of the radar cube, to obtain the signal content in frequency across all virtual channels, which is representing the reflections from different ranges called range profiles. The angle of arrival is obtained by exploiting a high-resolution, minimum variance distortion-less response (MVDR) Capon filter on each range bin [19]. Also, the stationary clutters on each range are removed since the interest is in finding the moving targets.
Then, the constant false alarm rate (CFAR) processor yields a cluster of points corresponding to the moving areas of the targets (see Fig. 1a). Also, after clustering, the centre point of each cluster is taken to recover the slow-time respiration signal.
After CFAR, clustering should be able to assign a group of points to one subject. So, the clustering process must determine the number of targets automatically, the mass centre of each target, and should be computationally efficient. Among many clustering algorithms, we found mean shift is best fitted to our application with reasonable computation burden 1 . Mean shift requires to know the typical minimum distance of targets, which we set it to 50 cm.
After clustering and finding the centre mass of each target cluster, the waveform of the target extracts from the target range and angle. In fact, the radar cube in Fig. 1b contains the breathing signal of the targets not only at different ranges but also on different azimuths. To extract each target signal individually, the following angular matched filter is applied: where s ij (t s ) is the radar response in slow time at the i'th range bin andθ j angle. Also, w(θ j ) is the steering vector for the representative point of the target cluster, andx is the received vector across all virtual channels after range FFT. In fact,θ j is angle of the cluster centroid, and w(θ j ) is formed based on the location of MIMO virtual array. For a linear uniform array with inter-element spacing of d, that is: For our radar, the number of virtual channels is N = 8, and d = λ/2. The range of centroids are used to extract the complex breathing waveform as an input to the optimum harmonic estimator filter.

C. Sleep Position Detection
Knowing the sleep position is important to change the sleep position of patients more often as a relief to the bed ulcers. Also, sleep apnea patients need to change their position preceding the obstruction moments. Conventionally, the position is given by a wearable device, a sensor under the bed, or a camera. None of them provides flexibility and freedom to the subject as the radar does. So, the radar advises the patient's supervisor about the sleep position history, and the supervisor helps to change the position of the patient has stayed on a position for a long time.
Association and CFAR together provide a collection of points for each target over time. These points show the spatial distribution of the targets. The coordinates of the points is with respect to the radar look direction. Since we use the radar above beds, see Fig. 2a, so "x" is along the width of beds and "y" is perpendicular to the ground surface (see Fig. 2b). There are features in the points that are different if the target sleeps on the back or on the sides. We define features from CFAR points of each subject rather than the radar cube maps, such as range-angle map, to classify each target independently. Our feature selection depends on the target point distributions in space. However, we have observed that Capon CFAR output points are very close in both sleep positions. Unlike slightly different xy point distributions for the two sleep positions, the spatial gradient of the Capon maps is completely different since intuitively the motions around the shoulders drops quicker than around the chest. So, we take a Capon map and apply derivative in both x and y directions laying out two new maps for the x and y derivatives. These derivatives have the following relationship to the range and the angle: dx = sin(θ) dr + r cos(θ) dθ dy = cos(θ) dr − r sin(θ) dθ CFAR points of each target and the corresponding derivatives and the power are features for the target (see Fig. 1b). So, we have space-time-frequency features making a 5 dimensional space of {x, y, dx, dy, power}. The feature space suggests that in 5-dimensional space, there is a nonlinear decision boundary for classifying the sleep positions. The data is split into training and test sets with 20% for test set. We used a grid search for hyper parameter optimization of SVM using 5fold cross validation with the following search space for each parameter.

III. EXPERIMENTS
We performed experiments in a bedroom with two beds, as shown in Fig. 2a. The subjects were allowed to have a slight motion during sleeping like "sliding the blanket" or "moving hands and legs", but their faces were up toward the radar. The radar and system parameters are listed in the table in 1a. We will show how the vital signs and sleep positions are obtained in the two experiments. We examined subjects for different sleep positions.

A. Vital signs detection
We conducted three rounds of experiments for two subjects, so, we have 6 subjects in total, who were breathing normally. During these experiments breathing and its waveform were estimated. In each round, two subjects were sleeping for about 20 minutes, and they were sleeping on their back. Fig. 2b shows the output map of CFAR in which the two targets are spotted with point clouds. Although the targets were almost on the same range to the radar, the use of MIMO radar distinguishes them in the angle. Specifically, the left and right subjects are in 1.81 and 1.76 meters to the radar, respectively. In Fig. 2a, the lateral separation of targets was 1 meter, which is enough for the radar angular resolution to resolve the two subjects at this range. At the beginning and at the end of the recording, when the subjects lay down on the bed and when they woke up, their motions are quite high such that the radar signal is highly distorted, and it is not reliable for breathing monitoring as illustrated in Fig. 2c and Fig. 2e. In contrast, in the middle of the test interval, they fell deeper into asleep, and the radar can detect their breathing, as shown in the two figures. The comparison with the reference sensor and the radar shows that the right subject had extra motions at around 130, 177, and 1079 seconds, and this happens for the left subject at around 84, 264, and 936 seconds, which are encircled in Fig. 2c and Fig. 2e.
Using optimum filter does not only enhance and separate periodic signals, but it provides an estimate of the signal waveform as well. As illustrated in Fig. 2d and Fig. 2f, the radar signal contains many harmonics due to reasons as we discussed before. However, the filter output is a noiseless waveform corresponding to a single frequency signal with the frequency of breathing rate, which is labeled in the figures with reconstructed signal. The magnitude of signals is related to the thorax displacement in a respiration cycle, which represents the lung volume change.
The system performance is evaluated based on two metrics: average error rate and root mean square error (RMSE). The former represents the error when the BR estimated by the radar is not closer than 3 bpm to the true value, and the latter gives the error standard deviation in bpm. To show the breathing detector for side sleeping, a patient slept on her side for about 20 minutes. The average of the metrics for all experiments are illustrated in Fig. 3. In all analysis, the moments with body motions are eliminated. The BR correct detection rate for left and right subjects sleeping on back are 99% and 94%, respectively. Also, the breathing rate analysis for the side sleeping is done for only one target. So, a single bar shows an even lower error rate than the back sleeping in Fig. 3, and the detection rate is 99% in this case. Overall, our system shows 97% accuracy in respiration rate analysis. In addition, RMSE values imply that the radar detection has maximum 1.7 bpm deviation over a long period.

B. Sleep position detection
In this study, we determine either the patient is sleeping on either right/left sides or on the back facing the radar. The data for training SVM machine learning classifier is collected from two experiments in which either the subjects were sleeping on their sides or on their backs. Each round was around 20 minutes. For each sleep position, the data length is 40 minutes per target. So, almost the sample size for back and side sleeping are the same.
As mentioned before, the derivative of the Capon map for each target is used as two features (dX and dY). These features are recorded for each target at the vital signs frame rate. Therefore, for every frame we have two targets and two samples for the training SVM. By the grid search, the best estimator is found with the parameters listed in Table I, and the confusion matrix shows at least 83% correct detection of the side or front sleeping (see Fig. 4).

IV. CONCLUSIONS
We proposed and developed a system for bedbound patients to monitor breathing and their sleep positions. The novel algorithm detects closely spaced sleeping subjects with a single radar and analyzes them individually. This advantage is due to using an array at the receiver whereas we achieve a higher number of target detection by virtually extending the number of receiver elements using MIMO radar. Therefore, instead of using multiple sensors per patient, which are having a single functionality of either monitoring breathing or sleep position, we employed a single radar to carry out those measurements. Replacing all the conventional sensors with only one sensor reduces the cost and power while maintaining reliable accuracy. The proposed system has two monitoring tasks: breathing and sleep position. For breathing, a spectral estimator was designed with 1 bpm resolution and an update rate of 1 second to estimate the breathing rate and its noiseless waveform. We found that the lateral distance of the subjects on the bed could be less than 1 meter (back-to-back), which is great for monitoring two people on the same bed. For sleep position detection, an SVM machine learning classifier with an appropriate set of features proposed to determine the side or back sleeping for each patient independently. Moreover, the bedroom experiments validated the system performance in a practical setting. The sleep position detector was successful more than 83%. Also, the breathing rate accuracy tested for two different sleep positions. For the back sleeping, the breathing rate was at least 94% of the time closer than 1 bpm to the reference sensor. This was 99% for side sleeping. In addition, considering all measurements, the maximum RMSE of breathing rate was 1.7 bpm. Since the random body motions are the main source of distortions, ignoring the moments with the extra motions is critical. This could be done automatically by a smart detector, which determines the degree of body movements based on a set of criteria such as Doppler power in different frequency regions, the area size with high Dopplers, and so on. Also, lung volumes are directly related to the respiration effort, which is critical for sleep apnea patients. For this, the exact lung volumes could be measured with a linear map between the radar and the actual values from the reference sensor.