黃毅 孫為軍 王丹雷 伍賢美 袁永浩
開(kāi)發(fā)設(shè)計(jì)
基于ResNet與BiLSTM的心電QRS波群檢測(cè)方法
黃毅 孫為軍 王丹雷 伍賢美 袁永浩
(廣東工業(yè)大學(xué),廣東 廣州 510006)
為快速準(zhǔn)確地檢測(cè)出心電信號(hào)中的QRS波群,提出一種基于殘差網(wǎng)絡(luò)(ResNet)與雙向長(zhǎng)短期記憶網(wǎng)絡(luò)(BiLSTM)的深度學(xué)習(xí)模型—ResBiLSTM,檢測(cè)QRS波群的起始點(diǎn)和終點(diǎn)。實(shí)驗(yàn)結(jié)果表明:相比于傳統(tǒng)的QRS波群檢測(cè)方法,ResBiLSTM提高了檢測(cè)效率,且具有較強(qiáng)的魯棒性,能夠準(zhǔn)確檢測(cè)不同形態(tài)的QRS波群。
心電信號(hào);QRS波群;殘差網(wǎng)絡(luò);雙向長(zhǎng)短期記憶網(wǎng)絡(luò)
QRS波群是心電信號(hào)最顯著和最重要的特征[1],通過(guò)檢測(cè)QRS波群可以獲取用于診斷的生物信息,如心率和呼吸信號(hào)[2]可用于心跳分類[3]等。目前,QRS波群檢測(cè)方法較多,傳統(tǒng)的QRS波群檢測(cè)方法有小波變換法[4-6]、希爾伯特變換法[7]和向量變換法[8]等。但這些檢測(cè)方法存在提取特征過(guò)于依賴經(jīng)驗(yàn)、模型參數(shù)固定和變換域計(jì)算量大等問(wèn)題,導(dǎo)致檢測(cè)算法魯棒性差、效率低。
近年來(lái),深度學(xué)習(xí)在計(jì)算機(jī)視覺(jué)和語(yǔ)音識(shí)別等領(lǐng)域應(yīng)用廣泛。卷積神經(jīng)網(wǎng)絡(luò)和循環(huán)神經(jīng)網(wǎng)絡(luò)也逐漸應(yīng)用于心電信號(hào)QRS波群的特征提取和檢測(cè)定位,這些神經(jīng)網(wǎng)絡(luò)可實(shí)現(xiàn)端到端的QRS波群檢測(cè),解決了傳統(tǒng)檢測(cè)方法模型參數(shù)固定和變換域計(jì)算量大等問(wèn)題。XIANG等[9]搭建兩級(jí)一維卷積神經(jīng)網(wǎng)絡(luò)模型,在MIT-BIH心率失常數(shù)據(jù)庫(kù)中獲得的靈敏度和陽(yáng)性預(yù)測(cè)值分別為99.77%和99.91%。YUEN等[10]搭建CNN-LSTM神經(jīng)網(wǎng)絡(luò)模型,在MIT-BIH噪聲壓力數(shù)據(jù)庫(kù)和歐洲ST-T噪聲壓力數(shù)據(jù)庫(kù)中都取得較好的QRS波群定位結(jié)果。CAI等[11]提出2種基于多拓展卷積塊的神經(jīng)網(wǎng)絡(luò)模型,在MIT-BIH心率失常數(shù)據(jù)庫(kù)中獲得最高靈敏度和最高陽(yáng)性預(yù)測(cè)值分別是99.95%和99.97%。雖然這些深度學(xué)習(xí)模型的QRS波群定位精度較高,但沒(méi)有檢測(cè)QRS波群的起始點(diǎn)和終點(diǎn),無(wú)法通過(guò)這些模型計(jì)算QRS波群寬度。而QRS波群寬度是心電圖臨床醫(yī)學(xué)診斷的重要參數(shù),常被用作心律失常的診斷依據(jù)。
本文提出一種基于殘差網(wǎng)絡(luò)(residual network, ResNet)與雙向長(zhǎng)短期記憶網(wǎng)絡(luò)(bidirectional long short-term memory, BiLSTM)的心電QRS波群起始點(diǎn)和終點(diǎn)檢測(cè)方法。該方法通過(guò)構(gòu)建ResBiLSTM進(jìn)行QRS波群的特征提取和檢測(cè),實(shí)現(xiàn)端到端的QRS波群檢測(cè),從而提高QRS波群的檢測(cè)效率。
QT數(shù)據(jù)庫(kù)[13]共有105條15 min的雙導(dǎo)聯(lián)心電信號(hào)記錄,每條心電信號(hào)記錄由1個(gè)頭文件、1個(gè)信號(hào)文件和9個(gè)注釋文件組成,采樣頻率為250 Hz。其中,record.q1c注釋文件對(duì)信號(hào)文件中的3623個(gè)QRS波群做了標(biāo)記。MARTíNEZ等[6]、DI MARCO等[14]、BOTE等[15]和王大雄等[16]在record.q1c文件標(biāo)記QRS波群數(shù)據(jù)上做QRS波群起始點(diǎn)和終點(diǎn)的檢測(cè)。record.pu0注釋文件對(duì)信號(hào)文件中222026個(gè)QRS波群做了標(biāo)記。本文在record.pu0文件標(biāo)記QRS波群數(shù)據(jù)上做QRS波群起始點(diǎn)和終點(diǎn)的檢測(cè),并從QT數(shù)據(jù)集中選擇92條心電信號(hào)記錄,其中64條用來(lái)訓(xùn)練模型,剩下28條用來(lái)測(cè)試模型。
在心電信號(hào)QRS波群起始點(diǎn)和終點(diǎn)檢測(cè)過(guò)程中,去除噪聲是關(guān)鍵一步。如果直接將帶有基線漂移、高頻肌電干擾等噪聲的心電信號(hào)輸入ResBiLSTM模型,受這些噪聲干擾的模型無(wú)法準(zhǔn)確學(xué)習(xí)QRS波群的有效特征,從而導(dǎo)致模型檢測(cè)QRS波群起始點(diǎn)和終點(diǎn)的準(zhǔn)確率下降。目前,一般采用帶通濾波器[17]或小波變換[18-19]去除心電信號(hào)噪聲。本文采用基于帶通巴特沃斯濾波器的前向后向?yàn)V波器[20]去除基線漂移和高頻肌電干擾噪聲。心電信號(hào)Sel30去噪前圖和去噪后圖分別如圖1、圖2所示。
圖1 sel30去噪前
圖2 Sel30去噪后
QRS波群標(biāo)簽編碼后,其圖形是一條幅度為1的脈沖波形,脈沖波形的上升沿和下降沿位置分別對(duì)應(yīng)QRS波群的起始點(diǎn)和終點(diǎn)。心電信號(hào)Sele0210的QRS波群標(biāo)簽編碼圖如圖3所示。
圖3 Sele0210的QRS波群標(biāo)簽編碼
受2015年HEK提出的ResNet[12]結(jié)構(gòu)啟發(fā),本文采用一維卷積神經(jīng)網(wǎng)絡(luò)構(gòu)建4層殘差網(wǎng)絡(luò)模塊;再結(jié)合一維卷積層,雙向長(zhǎng)短期記憶網(wǎng)絡(luò)和全連接層構(gòu)建ResBiLSTM模型用來(lái)檢測(cè)QRS波群起始點(diǎn)和終點(diǎn),模型結(jié)構(gòu)如圖4所示。
ResBiLSTM模型的第一層一維卷積層有4個(gè)卷積核,卷積核大小為11×1,激活函數(shù)是LeakyReLU。4層殘差網(wǎng)絡(luò)模塊里的一維卷積層有4個(gè)卷積核,卷積核大小為21×1,激活函數(shù)是LeakyReLU。BiLSTM層有128個(gè)神經(jīng)元,激活函數(shù)是ReLU。全連接層有2個(gè)神經(jīng)元,激活函數(shù)是Softmax。ResBiLSTM模型訓(xùn)練時(shí),選擇Adam作為訓(xùn)練優(yōu)化器,設(shè)置學(xué)習(xí)率為10-3,訓(xùn)練60個(gè)epoch,批量數(shù)據(jù)大小為128條切片數(shù)據(jù)。
最終模型得到采樣點(diǎn)屬于QRS波群的預(yù)測(cè)概率值是一條幅度為1的脈沖波形,脈沖波形的上升沿和下降沿位置就是QRS波群起始點(diǎn)和終點(diǎn)的檢測(cè)位置。
為驗(yàn)證本文提出的基于ResNet和BiLSTM的心電QRS波群起始點(diǎn)和終點(diǎn)檢測(cè)方法的有效性,用測(cè)試集數(shù)據(jù)測(cè)試ResBiLSTM模型。在QRS波群起始點(diǎn)和終點(diǎn)的檢測(cè)過(guò)程中,查找模型檢測(cè)的QRS波群起始點(diǎn)和終點(diǎn)附近150 ms[21]內(nèi)是否存在對(duì)應(yīng)注釋QRS波群的起始點(diǎn)和終點(diǎn)。
QRS波群基點(diǎn)包括起始點(diǎn)和終點(diǎn),評(píng)估其檢測(cè)方法有4個(gè)指標(biāo):敏感度()、陽(yáng)性預(yù)測(cè)值()、QRS波群基點(diǎn)位置的平均誤差()和QRS波群基點(diǎn)位置誤差標(biāo)準(zhǔn)差()。表示所有正樣本有多少被模型判為正樣本;表示模型判為正的所有樣本中有多少是真正的正樣本;±反映QRS波群基點(diǎn)的位置誤差數(shù)據(jù)分布。這4個(gè)指標(biāo)可以通過(guò)式(3)~式(6)計(jì)算。
QRS波群起始點(diǎn)和終點(diǎn)檢測(cè)如圖5所示。圖5中的(a)圖、(b)圖和(c)圖都是由上下子圖組成,其中上子圖是QRS波群標(biāo)簽圖,下子圖是QRS波群檢測(cè)圖。
圖5 (a) Sel808的正常QRS波群起始點(diǎn)和終點(diǎn)檢測(cè)
圖5 QRS波群起始點(diǎn)和終點(diǎn)檢測(cè)
由圖5可知:本文提出方法能準(zhǔn)確檢測(cè)正常QRS波群,倒置QRS波群和雙峰QRS波群。
在QT數(shù)據(jù)集上,本文提出的基于ResNet與BiLSTM的心電QRS波群起始點(diǎn)和終點(diǎn)檢測(cè)方法與其他檢測(cè)方法比較結(jié)果如表1所示。
表1 不同QRS波群起始點(diǎn)和終點(diǎn)檢測(cè)方法的比較結(jié)果
本文提出一種基于ResNet與BiLSTM的心電QRS波群起始點(diǎn)和終點(diǎn)檢測(cè)方法,解決了傳統(tǒng)QRS波群檢測(cè)方法存在提取特征過(guò)于依賴經(jīng)驗(yàn)、模型參數(shù)固定和變換域計(jì)算量大等問(wèn)題,具有檢測(cè)效率高、魯棒性強(qiáng)等特點(diǎn),適用于實(shí)時(shí)的QRS波群檢測(cè)。
[1] SHARMA T, SHARMA K. QRS complex detection in ECG signals using locally adaptive weighted total variation denoising[J]. Computers in Biology and Medicine, 2017,87: 187-199.
[2]LANGLEY P, BOWERS E J, MURRAY A. Principal component analysis as a tool for analyzing beat-to-beat changes in ECG features: application to ECG-derived respiration[J]. IEEE Transactions on Biomedical Engineering,2010,57(4):821-829.
[3] BANERJEE S, MITRA M. Application of cross wavelet transform for ECG pattern analysis and classification[J].IEEE Transactions on Instrumentation and Measurement,2014,63(2): 326-333.
[4] LI C , ZHENG C,TAI C. Detection of ECG characteristic points using wavelet transforms[J]. IEEE Transactions on Biomedical Engineering, 1995, 42(1):21-28.
[5] SAHAMBI J S, TANDON S N, BHATT R K P. Using wavelet transforms for ECG characterization. An on-line digital signal processing system[J]. IEEE Engineering in Medicine and Biology Magazine, 1997, 16(1):77-83.
[6] MARTíNEZ J P, ALMEIDA R, OLMOS S, et al. A wavelet-based ECG delineator: evaluation on standard databases[J]. IEEE Transactions on Biomedical Engineering, 2004, 51(4): 570-581.
[7] BENITEZ D, GAYDECKI P A, ZAIDI A, et al. The use of the Hilbert transform in ECG signal analysis[J]. Computers in Biology and Medicine, 2001, 31(5): 399-406.
[8] MARTINEZ A, ALCARAZ R, RIETA J J. Automatic electrocardiogram delineator based on the phasor transform of single lead recordings[C]//2010 Computing in Cardiology. IEEE, 2010: 987-990.
[9] XIANG Y , LIN Z , MENG J . Automatic QRS complex detection using two-level convolutional neural network[J]. Biomedical Engineering Online, 2018, 17(1):13.
[10] YUEN B, DONG X, LU T. Inter-patient CNN-LSTM for QRS complex detection in noisy ECG signals[J]. IEEE Access, 2019, 7: 169359-169370.
[11] CAI W , HU D . QRS complex detection using novel deep learning neural networks[J]. IEEE Access, 2020 (99):1-1.
[12] HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition, 2016: 770-778.
[13] LAGUNA P, MARK R G, GOLDBERG A, et al. A database for evaluation of algorithms for measurement of QT and other waveform intervals in the ECG[C]//Computers in cardiology 1997. IEEE, 1997: 673-676.
[14] DI MARCO L Y, CHIARI L. A wavelet-based ECG delineation algorithm for 32-bit integer online processing[J]. Biomedical Engineering Online, 2011, 10(1): 23.
[15] BOTE J M, RECAS J, RINCóN F, et al. A modular low-complexity ECG delineation algorithm for real-time embedded systems[J]. IEEE journal of Biomedical and Health Informatics, 2018, 22(2): 429-441.
[16] 王大雄,蔣云良,徐耕,等.基于LS估計(jì)檢測(cè)QRS波群寬度[J].中國(guó)生物醫(yī)學(xué)工程學(xué)報(bào),2010,29(3):321-325.
[17] PAN J, TOMPKINS W J. A real-time QRS detection algorithm[J]. IEEE Transactions on Biomedical Engineering, 1985,32(3): 230-236.
[18] KADAMBE S, MURRAY R, BOUDREAUX-BARTELS G F. Wavelet transform-based QRS complex detector[J]. IEEE Transactions on Biomedical Engineering, 1999, 46(7): 838-848.
[19] ELHANINE M, ABDELMOUNIM E, HADDADI R, et al. Electrocardiogram signal denoising using descrete wavelet transform[C]// 2014 International Conference on Multimedia Computing and Systems(ICMCS).IEEE, 2014.
[20] GUSTAFSSON F. Determining the initial states in forward-backward filtering[J]. IEEE Transactions on Signal Processing, 1996, 44(4): 988-992.
[21] Association for the Advancement of Medical Instrumentation. Testing and reporting performance results of cardiac rhythm and ST segment measurement algorithms[R]. ANSI/AAMI EC38, 1998, 1998.
[22] CSE Working Party. Recommendations for measurement standards in quantitative electrocardiography[J]. European Heart Journal, 1985, 6(10): 815-825.
QRS Complexes Detection Method of Electrocardiogram Signal Based on ResNet and BiLSTM
Huang Yi Sun Weijun Wang Danlei Wu Xianmei Yuan Yonghao
(Guangdong University of Technology, Guangzhou 510006, China)
In order to detect QRS complexes quickly and accurately, a deep learning model based on residual network (ResNet) and bidirectional long short-term memory network (BiLSTM) is proposed and the model is ResBiLSTM. The ResBiLSTM can detect the start and end points of QRS complexes. The experimental results show that compared with the traditional QRS complexes detection methods, ResBiLSTM not only improves the detection efficiency, but also has strong robustness ,which can accurately detect different forms of QRS complexes.
electrocardiogram; QRS complexes; residual network; bidirectional long short-term memory
TN911.7; TP183
A
1674-2605(2021)01-0005-06
10.3969/j.issn.1674-2605.2021.01.005
黃毅,男,1993年生,碩士研究生,主要研究方向:深度學(xué)習(xí),模式識(shí)別,生物信號(hào)處理。E-mail:1216621782@qq.com