Analysis of Abnormal Intra-QRS Potentials in Signal-Averaged Electrocardiograms Using a Radial Basis Function Neural Network
Abnormal intra-QRS potentials (AIQPs) are commonly observed in patients at high risk for ventricular tachycardia. We present a method for approximating a measured QRS complex using a non-linear neural network with all radial basis functions having the same smoothness. We extracted the high frequency, but low amplitude intra-QRS potentials using the approximation error to identify possible ventricular tachycardia. With a specified number of neurons, we performed an orthogonal least squares
... thm to determine the center of each Gaussian radial basis function. We found that the AIQP estimation error arising from part of the normal QRS complex could cause clinicians to misjudge patients with ventricular tachycardia. Our results also show that it is possible to correct this misjudgment by combining multiple AIQP parameters estimated using various spread parameters and numbers of neurons. Clinical trials demonstrate that higher AIQP-to-QRS ratios in the X, Y and Z leads are visible in patients with ventricular tachycardia than in normal subjects. A linear combination of 60 AIQP-to-QRS ratios can achieve 100% specificity, 90% sensitivity, and 95.8% total prediction accuracy for diagnosing ventricular tachycardia. Sensors 2016, 16, 1580 2 of 16 the delay in the conduction of electrical activity. Although a reentrant excitation event, caused by an arrhythmic substrate, is not accompanied by VLP in all instances, it is likely to appear as part of AIQP in a QRS complex [9,10]. As an alternative form of VLP, AIQP is able to completely reflect a reentrant excitation due to an arrhythmic substrate, and may improve the diagnostic performance of a signal-averaged electrocardiogram. Extracting AIQPs is a challenging task, since they are extremely weak notch and slur signals with abrupt changes in slope, embedded in a QRS complex. Moreover, AIQPs are essentially random signals that differ among VT patients. As presented by Gomis et al.  and Lander et al.  , AIQPs are estimated in a discrete-time cosine transform (DCT)-domain autoregressive moving average (ARMA) model as the residual signal, that is, the difference between a target QRS complex and a synthesized one. One remarkable advantage is that the AIQP of interest can be separated from a normal QRS complex by means of a low order ARMA model. This approach can achieve a total prediction accuracy (TPA) of between 63.6% and 75.8% for diagnosing VT patients, provided that either X, Y or Z lead AIQP parameters are employed. In contrast, a linear prediction model may be employed to extract the signals with abrupt slope change embedded in a QRS complex [13, 14] ; these are referred to as the unpredictable intra-QRS potentials (UIQPs). The analysis of UIQPs has been validated experimentally as an effective approach to identifying the presence of AIQPs, and accordingly, as a way to diagnose VT patients. Use of either X, Y or Z lead UIQP parameters provided a TPA between 76.4% and 83.3% in one study  , and a TPA between 72.2% and 79.2% in another  . A wavelet transform-based approach has been proposed by Tsutsumi et al.  and Yodogawa et al.  as a way to analyze the high frequency components of a QRS complex, and to diagnose patients with lethal ventricular arrhythmias. Specificity as high as 79.4%-93.8%, but sensitivity as low as 23.2%-37.2% were found in patients with lethal ventricular arrhythmias , whereas another study showed sensitivity up to 96%, but specificity of 64.3%, were seen among patients with VT or ventricular fibrillation  . As presented in a great number of previous studies, a skillful combination of AIQP and VLP parameters can further improve diagnostic accuracy for patients at high risk of ventricular arrhythmia, but improved accuracy of the AIQP estimates is required for clinical application. Previously, most AIQP estimates were made with linear models or by means of linear transformation. It may be better to model a QRS complex as a nonlinear signal, due to the elaborate working mechanism of a human heart. To estimate AIQPs, our previous study  employed a nonlinear radial basis function neural network (RBFNN), as commonly seen in the disciplines of function approximation  and data classification  . In this manner, a strong, slow varying, normal QRS complex is synthesized by an RBF neural network, and an approximation error is regarded as the weak, rapidly varying AIQPs. Short of clinical trials, a great amount of neurons are employed to build an RBF neural network, and AIQPs are modeled as white noise  . This work is an improved version of our prior research, applied to diagnosing patients with VT. Nevertheless, it is a non-trivial and challenging task to precisely extract the AIQP of interest from a QRS complex, because the approximation error may not merely include the wanted contribution from the AIQP, it may also comprise the unintended contribution from part of the normal QRS complex. One major problem is that a decrease in the approximation error may occur in the event that both contributions are out of phase, which could lead a clinician to misjudge the presence of AIQP. There is no way to further extract the AIQP estimation error caused by part of the normal QRS complex, since in practical situations, AIQPs are unpredictable and random signals. In this study, we conducted an in-depth investigation into the AIQP estimation error with AIQPs simulated as multiple noise sources, and found improved accuracy in clinical trials using linear combinations of the AIQP parameters. In short, the aim of this work is: (1) to develop a systematic RBF neural network-based approach for AIQP estimation from a measured QRS complex, as a high performance diagnostic tool for VT patients; (2) to evaluate the effects of AIQP estimation errors caused by part of the normal QRS complex; and (3) to improve the diagnostic accuracy by taking a linear combination of AIQP parameters.