Contactless Heartbeat Detection from CW-Doppler Radar using Windowed-Singular Spectrum Analysis* Yuki IWATA1 , Koichiro ISHIBASHI1 , Guanghao SUN1 , Manh Ha LUU2 , Trong Thanh HAN3 , Linh Trung NGUYEN2 and Trong Tuan DO3 and the frequency-domain approach have been developed to retrieve cardiac information from the demodulated signal[7][8][9][10] These approaches require significant measurement times and involve high computational costs Recently, we reported a time-domain approach using the analog band-pass filter (BPF) and the peak detection method for heartbeat detection in infection screening systems; this method requires a short measurement time[11] However, this study did not account for the influence of RBM noise Abstract— The continuous-wave Doppler radar measures the movement of a chest surface including of cardiac and breathing signals and the body movement The challenges associated with extracting cardiac information in the presence of respiration and body movement have not been addressed thus far This paper presents a novel method based on the windowed-singular spectrum analysis (WSSA) for solving this issue The algorithm consists of two processes: signal decomposition via WSSA followed by the reconstruction of decomposed heartbeat signals through convolution An experiment was conducted to collect chest signals in 212 people by Doppler radar In order to confirm the effect of reducing the large noise by the proposed method, we evaluated 136 signals that were considered to contain respiration body movements from the collected signals When comparing to the performance of a band-pass filter, the proposed analysis achieves improved beat count accuracy The results indicate its applicability to contactless heartbeat estimation under involving respiration and body movements Index Terms— Doppler radar, Vital signs, WSSA In this paper, we propose a novel time-domain approach for heartbeat detection, considering that the movements of the chest surface due to heartbeats are significantly smaller than those due to RBM and breathing This approach consists of windowed-singular spectrum analysis (WSSA) and template matching by using the decomposed signal related to heartbeat as a template signal We applied this method on the dataset measured using an off-the-shelf 24-GHz radar sensor and a simple analog-front-end architecture to validate the proposed method The rest of paper will be organized as follows Section II discribes the principle of vital signs detection using CW-Doppler radar Section III elaborates the WSSA baced heartbeat detection technique In Section IV, the effectiveness of the proposed algorithm is evaluated by the data collected in a experiment Finally, Section V summarizes the concluding remarks I I NTRODUCTION In recent years, the continuous-wave (CW) Doppler radar technique has been attracting attention owing to its application in contactless measurement of vital signs Heartbeat and respiration measurements via the CW-Doppler radar technique is widely applied when non-contact detection and privacy protection is required, such as infection screening in quarantine stations and sleep monitoring systems in home healthcare[1][2][3] The CW-Doppler radar measures the velocity of the movements of the chest surface consisting of cardiac and breathing signals as well as random body movement (RBM) noise However, RBM noise results in significant errors in the measurement due to its large amplitude and unsteadiness, especially in heartbeat detection[4] II BASIC F ORMULA OF CW-D OPPLER R ADAR Fig shows the basic structure of the Doppler radar for vital sign detection[4] The transmitter Tx transmits a continuous wave T (t) = AT cos[2π f t + ϕ (t)] toward the human body surface Where AT is the amplitude of the transmitted signal, f is the carrier frequency, and ϕ (t) is the phase noise According to the Doppler principle, the frequency of reflected wave changes depending on the body surface motion composed of respiration xr , heartbeat xh , and RBM xm Therefore, the receiver Rx receives the reflected wave R(t), which is expressed as Eq.(1): [ ] 4π d0 4π x(t) R(t) = AR cos 2π f t − − + ϕ (t − 2d0 c) (1) λ λ In previous studies, significant efforts have been devoted to solving this problem The arctangent as a demodulation method was proposed to account for the movement of large variable objects[5][6] Moreover, the blind source separation *This work was supported by JSPS KAKENHI Grant Number JP19H02385 Yuki IWATA, Koichiro ISHIBASHI and Guanghao SUN are with Guraduate School of Informatics and Engineering, The University of Electro-Communications(UEC), 1-5-1 Chofugaoka, Chofu, Tokyo, Japan Where AR is the amplitude of the received signal, λ is the wavelength, c is the speed of light, d0 is the initial distance between the Doppler radar and the body surface, and x(t) = xr + xh + xm is displacement of human body surface As shown in Fig 1, when the received signal R(t) is down-converted, two baseband signals are obtained One yuki.iwata.as@uec.ac.jp Manh Ha LUU and Linh Trung NGUYEN are with the University of Engineering and Technology, Vietnam National University(VNU-UET), 144, Xuan Thuy, Cau Giay, Hanoi, Vietnam Trong Thanh HAN and Trong Tuan DO are with the School of Electronics and Telecommunications, Hanoi University of Sience and Technology(HUST), 1, Dai Co Viet Road, Hanoi, Vietnam 978-1-7281-1990-8/20/$31.00 ©2020 IEEE 477 Authorized licensed use limited to: Cornell University Library Downloaded on September 10,2020 at 05:48:43 UTC from IEEE Xplore Restrictions apply 提案手法 Doppler radar is the in-phase signal BI (t), and the other is the quadrature phase signal BQ (t) ] [ 4π {x(t) + d0 } + θ + ∆ϕ (t) , (2) BI (t) = AI cos λ [ ] 4π {x(t) + d0 } + θ + ∆ϕ (t) (3) BQ (t) = AQ sin λ Arctangent Proposed Method Windowed-SSA Template Formation where AI is the amplitude of the in-phase signal, and AQ is the amplitude of quadrature phase signal, θ is the phase constant, and ∆ϕ (t) is phase noise Because these signals have a “null point” problem, detection of the variable object generally requires demodulation[5] This study adopted the arctangent, it is a strong method that employs two orthogonal signals to demodulate phase information and is expressed as follows: ( ) BQ (t) λ x(t) ≈ arctan (4) 4π BI (t) Heartbeat Enhancement Peak Detection 20 2: January 2020 The algorithm of the proposed method Fig Subsequently, the trajectory matrix is decomposed (√ using) λi , u i , v i singular value decomposition into eigentriplet as Here, we assume that AI and AQ are equal and θ and ∆ϕ (t) are small 𝑥𝑥𝑚𝑚 𝑥𝑥ℎ T(𝑡𝑡) Tx 𝑥𝑥𝑟𝑟 R(𝑡𝑡) Rx 0°90° BI (𝑡𝑡) 𝑑𝑑0 BQ (𝑡𝑡) Step 20sec Normalized Normalized 1.3 1.2 III P ROPOSED M2sec ETHOD 1.1 490 495 500 505 510 The diagram of the Time[sec] proposed method for heartbeat data PCA detection is shown in Measurement Fig The baseband signals BI and PC1 Firstly, the signal are demodulated by BQ form the input Step Reject information related the arctangent method to ⋮ extract phase ⋮ Convolution to the displacement of the chest surface Subsequently, the 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 demodulated signal is applied to the proposed method to ⋮ ⋮ enhance the heartbeat Finally, the peaks of the heartbeat are 𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 PC detected from the9 time series, and Step its 3intervals are used to Feature Vectors calculate the heart rate per minute Time[sec] -5 0.5 1.5 0.5 1.5 0.5 1.5 0.5 1.5 -2 i=1 i=1 A Decomposition by WSSA WSSA was developed to reduce the high computational load of singular spectrum analysis (SSA), which performs non-parametric signal decomposition of time series[12] In this step, WSSA is used to decompose the movement of chest surface signals obtained from the CW-Doppler radar into meaningful components, such as body movement, breathing, and heartbeat Firstly, WSSA forms a trajectory matrix Y ∈ Rm×n from the time series of the demodulated signal x = {x1 , x2 , , xmn } as x1 x2 · · · xn xn+1 xn+2 · · · x2n Y = (5) ··· (6) In this step, we consider enhancing the heartbeat in the measurement signal using the eigentriplet belonging to group I2 However, in the diagonal averaging reconstruction method, which is used in conventional SSA and WSSA, the noises that are included in the eigentriplet of group I2 will appear In addition, it is difficult to identify the eigentriplet in group I2 that represents a heartbeat Therefore, we employ template matching for the reconstruction process Template matching is a method to emphasize the target signal expressed as 0.5 -0.5 x(m−1)n+2 √ λi u i v ′i , B Heartbeat Enhancement by Template Matching -1 x(m−1)n+1 d √ where λi , u i and v i are singular values, left and right singular vectors, respectively, of the ith eigentriplet d is the rank of the trajectory matrix Thereafter, the i = 1, , d is divided into two groups defiened as I1 = {1, , p} and I2 = p {p + 1, , d} Here, p is determined as ∑i=1 λi / ∑di=1 λi = ξ , where ξ is the thresthold In this study, we used n = 200 and ξ = 0.9 Where n and ξ were empirically determined Using this operation, the eigen triplet belonging to the group I1 and I2 can be expressed as components with a large amplitude (i.e., signals due to the body movement and respiration), and small amplitude (i.e., signals due to the heartbeat and small noise like a white noise), respectively Fig 1: Fundamental mechanism of CW-Doppler radar Step d Y = ∑Y i = ∑ 24GHz Proposed method 畳み込み再構成型-WSS (Convolutional Reconstructed-WSSA CR-WSSA) x R [n] = x [n] ⊗ h ∗ [−n] , (7) where x R is the time series of reconstructed signal, and h ∗ [−n] is the template signal, and ⊗ is the convolution operator Here, the template signal is composed of right singular vectors, as expressed in Eq.(8): h=∑ xmn I2 vi , ∥vvi ∥2 478 Authorized licensed use limited to: Cornell University Library Downloaded on September 10,2020 at 05:48:43 UTC from IEEE Xplore Restrictions apply (8) Phase[rad] where ∥ · ∥ is the norm operator By configuring the template with the sum of the singular vectors corresponding to the low singular values, we attempt to reduce the difficulty of determining the threshold IV E XPERIMENT A Experimental Conditions and Environment 〇：Heartbeat 10 〇〇〇 10 〇 〇 〇 〇 〇 〇〇 〇 〇 -5 00 10 Phase[rad] Phase[rad] 15 20 25 30 10 (a) 15 20 25 30 Corrupted by body movement 10 15 20 25 20 25 30 Time[sec] 10 15 Time[s] 30 (b) 25 15 10 40 -4 42 -3 36 -3 -2 30 -2 24 18 -1 12 0- (b) 20 PPG 80 BPF This work 60 BPF This work 6- Number of subjects 100 20 Cumulative frequency (%) Fig 4: Two types of radar signals classified via visual inspection (a) Example of good quality data, (b) Example of bad quality data Radar Error[bpm] Fig 3: (a) Experimental environment (b) Analog-front-end circuit (a) 25 B Experimental Results 80 BPF This work 60 BPF This work 15 10 40 39 36 - 33 30 - 24 - 18 - 12 - 0- 27 21 15 20 The proposed method was supposed to apply to signals containing large noise (i.e., body movement and respiration) For this reason, the dataset was divided into a good quality dataset and a bad quality dataset, based on visual inspection Fig 4a, and 4b presented an example of the radar signal in each dataset As shown in Fig 4a, the presence of small beat signals appearing at intervals of approximately s can be confirmed throughout and form the Good quality samples From Fig 4b, it is evident that the respiration signal is present in the time series signal; however, none of the signals resemble heart rates In addition, the radar signal is broken in the range of 17–30 s due to changes in body motion, and the presence cannot be confirmed along with breathing and heartbeat Radar signals with such shapes were classified as Bad quality samples 6- Number of subjects 100 20 Cumulative frequency (%) (a) 〇 Time[sec] 10 -10 -20 -10 -20 〇〇 〇 〇 〇〇 〇 〇〇 〇 〇〇〇 〇〇 Time[s] -5 10 For evaluating the proposed method, the displacement of the chest surfaces of 212 subjects (with an average age of 23 years) were measured for 30 s Fig 3a depicts the experimental environment The CW-Doppler radar was placed in front of a seated subject In this experiment, the baseband signals were produced by a 24 GHz CW-Doppler radar (NJR4262J, New-JRC, Tokyo, Japan), and the signals were passed through an analog-front-end (with a bandwidth of 0.5–50 Hz), as shown in Fig 3b The output signals were recorded using a 14-bit analog-to-digital converter (National Instruments, USB-6008) The sampling frequency was 100 Hz The measurement flow was controlled using a laptop PC with LABVIEW installed Moreover, the pulse wave signals were collected using the photoplethysmography (PPG) sensor as the grand trues of a heartbeat This study was approved by the Ethics Committee of the University of Electro-Communications 〇 〇 〇 〇 Error[bpm] (b) Fig 5: Comparison with two types of signal processings (a) Histogram showing the error of heart rate in good quality dataset, (b) Histogram showing the error of heart rate in bad quality dataset 479 Authorized licensed use limited to: Cornell University Library Downloaded on September 10,2020 at 05:48:43 UTC from IEEE Xplore Restrictions apply The divided datasets were analyzed using the conventional method, including [11] (band-pass filter and peak detection) and the proposed method, to compare the effects of heartbeat extraction Here, as parameters of the proposed method, the threshold value was set to 0.9 and the division interval was set to s (200 samples) This analysis was implemented in MATLAB Fig 5a and 5b show the frequency distribution representing errors of the heart rate estimated by the proposed method (shown by the gray bar) and the conventional method (shown by the white bar), when applied to Good and Bad quality dataset, respectively The dotted and solid lines show the cumulative distribution of heart rate errors in each method R EFERENCES [1] T Ohata, K Ishibashi and G Sun, Non-Contact Blood Pressure Measurement Scheme Using Doppler Radar, 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Berlin, Germany, 2019, pp 778-781 [2] K Higashi, G Sun and K Ishibashi, Precise Heart Rate Measurement Using Non-contact Doppler Radar Assisted by Machine-Learning-Based Sleep Posture Estimation, 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Berlin, Germany, 2019, pp 788-791 [3] M Kobayashi, G Sun, T Shinba, T Matsui and T Kirimoto, Simple and objective screening of major depressive disorder by heart rate variability analysis during paced respiration and mental task conditions, 2017 39th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Seogwipo, 2017, pp 1316-1319 [4] Q Lv et al., Doppler Vital Signs Detection in the Presence of Large-Scale Random Body Movements, in IEEE Transactions on Microwave Theory and Techniques, vol 66, no 9, pp 4261-4270, Sept 2018 [5] W Massagram, N M Hafner, B Park, V M Lubecke, A Host-Madsen and O Boric-Lubecke, Feasibility of Heart Rate Variability Measurement from Quadrature Doppler Radar Using Arctangent Demodulation with DC Offset Compensation, 2007 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Lyon, 2007, pp 1643-1646 [6] B Park, O Boric-Lubecke and V M Lubecke, ”Arctangent Demodulation With DC Offset Compensation in Quadrature Doppler Radar Receiver Systems,” in IEEE Transactions on Microwave Theory and Techniques, vol 55, no 5, pp 1073-1079, May 2007 [7] J Tu and J Lin, ”Fast Acquisition of Heart Rate in Noncontact Vital Sign Radar Measurement Using Time-Window-Variation Technique,” in IEEE Transactions on Instrumentation and Measurement, vol 65, no 1, pp 112-122, Jan 2016 [8] E Mogi and T Ohtsuki, ”Heartbeat detection with Doppler radar based on spectrogram,” 2017 IEEE International Conference on Communications (ICC), Paris, 2017, pp 1-6 Magnetics Japan, 1982, p 301] [9] T Zhangi, G Valerio, J Sarrazin and D Istrate, ”Wavelet-based analysis of 60 GHz Doppler radar for non-stationary vital sign monitoring,” 2017 11th European Conference on Antennas and Propagation (EUCAP), Paris, 2017, pp 1876-1877 [10] A Dell’Aversano, A Natale, A Buonanno and R Solimene, ”Through the Wall Breathing Detection by Means of a Doppler Radar and MUSIC Algorithm,” in IEEE Sensors Letters, vol 1, no 3, pp 1-4, June 2017, Art no 3500904 [11] X Yang, G Sun and K Ishibashi, Non-contact acquisition of respiration and heart rates using Doppler radar with time domain peak-detection algorithm, 2017 39th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Seogwipo, 2017, pp 2847-2850 [12] R Rekapalli, R Tiwari, Windowed SSA (Singular Spectral Analysis) for Geophysical Time Series Analysis, Journal of Geological Resource and Engineering, 2014, pp 167-173 For the proposal of method applied to the signals in the good quality dataset, the cumulative frequency presented in Fig 5a, shows that the number of subjects with heart rate errors of − bpm was less than those when using the conventional method In particular, the number of subjects with a heart rate error of less than bpm decreased from 22 to 13 In contrast, for the proposed method applied to the signals in the bad quality dataset, the cumulative frequency depicted in Fig 5b indicates an increase in the number of subjects with heart rate error of − 12 s Furthermore, the number of subjects with heart rate errors of less than bpm and bpm increased from to 19 and 25 to 36, respectively These results indicate that applying the proposed method to the bad quality dataset is effective for heartbeat detection V C ONCLUSION We proposed a novel algorithm that is capable of reducing large-amplitude noises to detect heartbeats from the CW-Doppler radar signals, under respiration and RBM noise We investigated the influence of the quality of the dataset on the error compared to the grand-truth of the estimated heart rate, to validate the proposed method In the experiment, the displacement of the chest surface produced by the CW-Doppler radar and the pulse wave measured by PPG for 212 subjects had been recorded for 30 s The dataset was divided into two groups: good and bad quality datasets based on whether or not large noises were superimposed, based on visual inspection The number of subjects with a heart rate error of − 12 bpm increased when using the proposed method on the bad dataset Thus, we conclude that the proposed method exhibits the potential for application as a contactless heartbeat detection method that accounts for respiration and body movement ACKNOWLEDGEMENT I am grateful to the students of VNU-UET and HUST for their assistance with the experiments in Vietnam We also thank X Yang for designing and developing the analog-front-end attached to the CW-Doppler radar 480 Authorized licensed use limited to: Cornell University Library Downloaded on September 10,2020 at 05:48:43 UTC from IEEE Xplore Restrictions apply ... Sign Radar Measurement Using Time-Window-Variation Technique,” in IEEE Transactions on Instrumentation and Measurement, vol 65, no 1, pp 112-122, Jan 2016 [8] E Mogi and T Ohtsuki, ? ?Heartbeat detection. .. bpm increased when using the proposed method on the bad dataset Thus, we conclude that the proposed method exhibits the potential for application as a contactless heartbeat detection method that... signals obtained from the CW-Doppler radar into meaningful components, such as body movement, breathing, and heartbeat Firstly, WSSA forms a trajectory matrix Y ∈ Rm×n from the time series of the demodulated