Resampling for Reliable Evaluation and Improved Performance Inautomatic Detection of electrode Misplacements In Ecg: Studies Based On Limb-Lead Misplacements

It is estimated that up to 4% of all ECGs are collected with misplaced electrode cables. Electrode misplacement can significantly alter the ECG and thus lead to wrong diagnoses. We developed an algorithm for automatic detection of electrode misplacement and evaluated its performance under consideration of a realistic misplacement rate. The algorithm consists of two discrimination steps. Step 1 determines whether an ECG has been correctly collected. If the prediction of Step 1 is misplacement, Step 2 then determines the type of misplacement. Previous algorithms reported their performances on a database including equal number of misplaced and correct records. We showed that these algorithms may have high false positive rates on misplacement prediction. Resampling was conducted at both the training and test sets for faithful evaluation and for optimal performance. The performance of Step 1 is 99.6% accuracy, 91.7% sensitivity, 99.9% specificity, 98.9% positive predictive value and 99.7% negative predictive value. The overall accuracy of Step 2 is 90.4%. Introduction 2Department of Family Medicine 3Division of Cardiology, University of Tennessee Medical Center, Knoxville, TN 37920 J Biom Tech Res Vol. 1. Issue. 1. 6000103 Figure 1: Examples of selected electrode placements: (a) the correct ECG; (b) the reversal between LA and RA electrodes; and (c) a misplacement, where RA takes the position of LA, LA takes the position of LL, and LL takes the position of RA. Figure adapted from Xia et al. (2012a). The reversal between the LA and RA electrodes is the most common electrode placement error, and many ECG recorders have builtin logic to recognize this problem [6]. Early algorithms use rule-based criteria to detect the reversal between LA and RA [7]. For example, the Marquette program suspects LA-RA interchange “if the QRS axis is between 90 and 270 degrees and the P axis is between 90 and 210 degrees” [8]. While the Marquette program yields accurate results (93.91% sensitivity and 100% specificity) for ECGs with clear P waves, the results are disappointing (39.34% sensitivity and 100% specificity) for ECGs without P waves [9]. A few other algorithms have been developed for different types of electrode misplacement [10-16]. The recent review by Batchvarov et al. [17] provides a systematic review and comparison of various algorithms. Among the existing algorithms, the work of Heden et al [12] and that of Kors and Herpen [15] present the highest classification accuracies. Heden et al. were the first to use artificial neural networks for detection of electrode misplacement. They used various ECG waves and intervals as features. When classifying the reversal of left arm electrode and right arm electrode, they obtained a specificity of 99.95% and sensitivity of 99.11% when the P wave was present and a specificity of 99.92% and sensitivity of 94.5% when the P wave was absent. They also obtained a specificity of 57.6% and sensitivity of 99.97% for the reversal of left arm and left leg electrodes and a specificity of 71.76% and sensitivity of 99.92% on average for the precordial lead reversals.Kors and Herpen were the first to use intrinsic relationships between different leads of the ECG as features in detection of cable reversals. They investigated 14 types of misplacement including the 5 types of limb cable reversals and 9 types of precordial cable reversals. On a set of 3,305 ECGs, they demonstrated excellent specificities (>=99.5%) for all 14 types of misplacement and very good sensitivities (>=93%) for all interchanges except the LA/LL reversal (sensitivity = 17.9%, specificity = 99.5%). In a recent study, Xia et al. [18] reviewed and compared the algorithm of Heden et al. and that of Kors and Herpenusing 3 different databases from Physionet [19]. The work of Xia et al. shows that existing algorithms work well on records of high signal quality and without severe distortions, but the performance is less satisfactory on records with noise and arrhythmias. problems, where each type of misplacement is examined against the correct ECG. As such, for a given ECG, whose type of placement is not known, it is not clear how accurately this ECG can be labeled. In contrast, this article considers a more general multi-class classification problem. We first determine whether the ECG has been collected with misplaced electrodes. Then, for a misplaced ECG, we determine what the type of misplacement is. Moreover, performances of previous algorithms were tested using test data sets that contain the same number of correct and misplaced ECGs. Recall that misplacements occur in 4% of all ECGs in reality. Thus, performance evaluation based on balanced data sets may be misleading. This article aims to overcome these difficulties. The paper is organized as follows. Section 2 describes various layouts of ECG electrodes and the methods to derive the different electrode misplacements. Section 3 describes the databases used in this work. Section 4provides detailed explanations on the methods. The results are presented in Section 5. Finally, the conclusion and discussion are presented in Section 6. In theory, the 10 electrode cables in the 12-lead ECG have more than 3.6 million permutations, leading to millions of possible reversals. However, reversals of precordial leads are extremely unlikely to occur and researchers usually consider just a few types of misplacements of precordial leads [17]. In this article, we focused our attention on misplacements involving only limb electrodes. Layout of ECG Electrodes and Derivation of Misplaced ECGs The 4 limb cables have 24 different permutations of placement on the 4 limbs. In the standard ECG, RL and LL electrodes are placed at the ankles. Therefore, reversing the RL and LL electrodes will not change the ECG since the potential difference between the two ankles is essentially zero. As a result, the 24 different types of electrode placement yield 12 different ECGs. Table 1shows the 11 ECGs from limb lead reversals compared to the correct ECG. Note that all cable reversals are derived from operations on the correct electrode placement. Moreover, certain types of misplacement are obtained using two sequential operations. For example, the type 7 placement involves the interchange between LA and RL followed by the interchange between LA and LL. In theory, the last 6 rows of Table 1are trivial to detect since one of the leads in these ECGs is a flat line since one of the arm electrodes is placed on the right leg. See Xia et al. [18] for more detailed description on limb lead reversals. The review article of Batchvarov et al. [17] discusses rules to identify common cable interchanges. J Biom Tech Res Vol. 1. Issue. 1. 6000103 Page 2 of 9 Citation: Xiaopeng Zhao, Henian Xia, Irfan Asif, Dale C Wortham, Elena G Tolkacheva (2014) Resampling for Reliable Evaluation and Improved Performance Inautomatic Detection Ofelectrode Misplacements In Ecg: Studies Based On Limb-Lead Misplacements. J Biom Tech Res 1(1): 9. The existing algorithms consider binary classification Table 1: ECGs from various types of misplacement compared to the correct ECG.