[雙語翻譯]--外文翻譯--經(jīng)驗?zāi)Ｊ椒纸夂徒y(tǒng)計學方法下的動態(tài)（holter）心電圖運動偽影檢測 （原文）

1、IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 59, NO. 6, JUNE 2012 1499Automatic Motion and Noise Artifact Detectionin Holter ECG Data Using Empirical Mode Decomposition and Statistical ApproachesJinseok Lee*, Member

2、, IEEE, David D. McManus, Sneh Merchant, and Ki H. Chon, Senior Member, IEEEAbstract—We present a real-time method for the detection ofmotion and noise (MN) artifacts, which frequently interferes with accurate rhythm ass

3、essment when ECG signals are collected from Holter monitors. Our MN artifact detection approach involves two stages. The first stage involves the use of the first-order intrinsic mode function (F-IMF) from the empirical

4、mode decomposition to isolate the artifacts’ dynamics as they are largely concentrated in the higher frequencies. The second stage of our approach uses three statistical measures on the F-IMF time series to look for char

5、acteristics of randomness and variability, which are hallmark signatures of MN artifacts: the Shannon entropy, mean, and vari- ance. We then use the receiver–operator characteristics curve on Holter data from 15 healthy

6、subjects to derive threshold values associated with these statistical measures to separate between the clean and MN artifacts’ data segments. With threshold values de- rived from 15 training data sets, we tested our algo

7、rithms on 30 additional healthy subjects. Our results show that our algorithms are able to detect the presence of MN artifacts with sensitivity and specificity of 96.63% and 94.73%, respectively. In addition, when we app

8、lied our previously developed algorithm for atrial fibrilla- tion (AF) detection on those segments that have been labeled to be free from MN artifacts, the specificity increased from 73.66% to 85.04% without loss of sens

9、itivity (74.48%–74.62%) on six subjects diagnosed with AF. Finally, the computation time was less than 0.2 s using a MATLAB code, indicating that real-time application of the algorithms is possible for Holter monitoring.

10、Index Terms—Atrial fibrillation (AF), empirical mode decom-position (EMD), Holter recording, motion and noise (MN) artifact detection, statistical method.I. INTRODUCTIONWE have recently developed an algorithm for accurat

11、e and real-time detection of atrial fibrillation (AF) that iswell- suited for continuous ECG monitoring applications [1].Manuscript received January 21, 2011; revised June 27, 2011 and October25, 2011; accepted November

12、5, 2011. Date of publication November 10, 2011; date of current version May 18, 2012. This work was funded in part by the Office of Naval Research work unit under Grant N00014-08-1-0244. Asterisk indicates corresponding

13、author.*J. Lee is with the Department of Biomedical Engineering, Worcester Poly-technic Institute, MA 01609 USA (e-mail: jinseok@wpi.edu).D. D. McManus is with the Cardiology Division, Departments of Medicineand Quantita

14、tive Health Sciences, University of Massachusetts Medical Center, Worcester, MA 01605 USA (e-mail: mcmanusd@ummhc.org).S. Merchant is with the Scottcare Corporation, Cleveland, OH 44135 USA(e-mail: smerchant@scottcare.co

15、m).K. H. Chon is with the Department of Biomedical Engineering, WorcesterPolytechnic Institute, MA 01609 USA (e-mail: kichon@wpi.edu).Color versions of one or more of the figures in this paper are available onlineat http

16、://ieeexplore.ieee.org.Digital Object Identifier 10.1109/TBME.2011.2175729Use of ECG monitors (e.g., Holter monitors) is common in the diagnosis and management of patients with, or at risk for, AF, given the paroxysmal,

17、short-lived, and frequently asymptomatic nature of this serious arrhythmia. Monitoring for AF is impor- tant because, despite often being paroxysmal and associated with minimal or no symptoms, AF is associated with sever

18、e ad- verse health consequences, including stroke, heart failure, and death [2]. Our test of accuracy of the AF algorithm was per- formed on noise-removed test databases, which also consisted of Holter recordings. Certai

19、nly, motion and noise (MN) artifacts are significant during Holter recordings and can lead to false de- tections of AF. Clinicians have cited MN artifacts in ambulatory monitoring devices as the most common cause of fals

20、e alarms, loss of signal, and inaccurate readings [3], [4].Previous computational efforts have largely relied on MN arti-fact removal, and some of the popular methods include linear fil- tering [5], adaptive filtering [6

21、], [7], wavelet denoising [8]–[10], and Bayesian filtering methods [11]. One main disadvantage of the adaptive filtering methods is that they require a reference sig- nal, which is presumed to be correlated in some way w

22、ith the MN artifacts. For mitigating this limitation, use of accelerometers to obtain a reference signal has resulted in some success [12], [13]; however, this approach has not been applied to Holter monitors. The wavele

23、t denoising approach attempts to separate clean and noisy wavelet coefficients, but it can be difficult to use since it requires identification of the location of each ECG morphology including the P and T waves [8]–[10].

24、 Bayesian filtering requires estimation of optimal parameters using any variant of Kalman filtering methods: extended Kalman filter, extended Kalman smoother, or unscented Kalman filter [11]. The main disadvantage of the

25、 Bayesian filtering approach is the improper assumption that noise has an additive Gaussian prob- ability density function. Further, the method requires R-peak locations for each cycle of ECG data.While the aforementione

26、d signal processing approaches havebeen applied, they are not appropriate, and consequently MN artifacts remain a key obstacle to the accurate detection of AF and atrial flutter, which is an equally problematic arrhythmi

27、a. A novel method to separate clean ECG portions from segments with MN artifacts in real time is urgently needed for more ac- curate diagnosis and treatment of clinically important atrial ar- rhythmias. For our paper, th

28、e aim is to detect the presence of MN artifacts; for Holter applications, there are a sufficient num- ber of clean segments in each recording that MN-contaminated segments can be discarded, thereby increasing the specifi

29、city of AF identification. Moreover, our AF detection algorithm is0018-9294/$31.00 © 2012 IEEELEE et al.: AUTOMATIC MOTION AND NOISE ARTIFACT DETECTION IN HOLTER ECG DATA 1501Fig. 2. Squared-IMF based on clean and n

30、oisy ECG signal. (a) Clean ECG segment. (b) Noisy ECG segment.Fig. 3. Simplified algorithm for MN artifact detection in an ECG segment byusing EMD and three statistical techniques.Fig. 1(a)] and noisy signals [see (Fig.

31、1(b)]. As shown in Fig. 2, the peak amplitudes of the clean signal [see Fig. 2(a)] are an order of magnitude higher than those of the MN-corrupted signal [see Fig. 2(b)], indicating that a threshold value can be derived

32、to separate between the two types of signals.With a normalized squared IMF, we determine the optimumlow noise level threshold (LNLT) value and define it as THLNLT. For each THLNLT value starting from 0 to 1 at an increme

33、nt of 0.05, we investigate the following three statistical indices: Shannon entropy to characterize randomness, a mean value to quantify LNLT level, and variance to quantify variability. If all values of Shannon entropy,

34、 mean, and variance are higher than threshold values of THent, THmean, and THvar, we declare the segment to be a noise-corrupted segment. The overall algorithm is summarized in Fig. 3. Once THLNLT and the thresholds for

35、maximum sensitivity and specificity are determined for each of the three statistical values (THent, THmean, and THvar) using the receiver–operator characteristic curve analysis on the data, as described in Section II-B,

36、no further heuristic tuning for the threshold values is required. We also investigated the optimum segment length (Lseg) for maximum sensitivity and specificity along with computational complexity.B. Data Acquisition I:

37、Data Collection and Determination of Optimal Threshold ValuesWe collected 5-lead ECG Holter recordings (Scottcare Corpo-ration) from 15 healthy subjects. Data were acquired at 180 Hzwith 10-bit resolution for 24 h. None

38、of the subjects had clini- cally apparent cardiovascular disease. The 15 healthy subjects comprised 8 females and 7 males of age 31.7 ± 3.4 years. During Holter recording, each subject was asked to perform routine d

39、aily activities. Among the acquired data, we collected 144 10-s noisy segments, where R-peaks were not clearly rec- ognizable due to MN artifacts. Along with the noisy segments, we collected 144 10-s clean segments, wher

40、e RR intervals were clearly discernible. Note that the decision to deem a segment noise corrupted or clean was based on the criterion of whether or not the R-peaks of the ECG waveforms were recognizable to the eye.For th

41、e selection of the optimal threshold set consisting ofTHLNLT, THent, THmean, and THvar, we searched every pos- sible combination among the 4-D vectors with the following interval increments:1) THLNLT varied from 0 to 1 a

42、t intervals of 0.05; 2) THent varied from 0 to 1 at intervals of 0.0001; 3) THmean varied from 0 to 1 at intervals of 0.0001; 4) THrmssd varied from 0 to 0.01 at intervals of 0.00001. The optimal threshold was determined

43、 according to a combi-nation of the four threshold values that provided the best accu- racy. The accuracy was calculated as follows:Accuracy = TP + TNTP + TN + FP + FN (1)where TP, TN, FP and FN are true positives, true

44、negatives, false positives, and false negatives, respectively. With the data length, Lseg = 5 s, we found the accuracy of 0.9688, and the sensitivity and specificity values of 0.9549 and 0.9792, respectively.1) Optimal D

45、ata Length and Computational Time: To deter-mine the optimum data length Lseg for MN artifacts detection, we repeated the aforementioned procedure with a segment size varying from 1 to 10 s at an increment of 1 s. Based

46、on each Lseg (1–10 s), we obtained the optimal parameters (e.g., 10 sets of threshold sets) and plotted the accuracy according to Lseg, as shown in Fig. 4(a). The accuracy increased when Lseg in- creased, but the rate of

47、 increase declined when Lseg was equal to or greater than 5 s. In addition, as shown in Fig. 4(b), the computation time for a clean segment linearly increased with the length of data segments. However, the computation ti

48、me for noisy segments dramatically increased especially when the segment length exceeded 6 s, as shown in Fig. 4(c). Taking into account the computational complexity, we chose the op- timum Lseg = 5 s. Note that the comp