INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

A SUBSTANTIAL BODY OF CLINICAL and experimental evidence indicates that T-wave alternans, a beat-to-beat fluctuation in the amplitude of that waveform, is fundamentally linked to vulnerability to ventricular fibrillation (VF) (2, 19, 24, 25, 26, 34, 37). In the experimental laboratory, it has been shown that T-wave alternans magnitude tracks vulnerability to ventricular tachyarrhythmias under diverse physiological and pharmacological interventions (2, 19, 24, 25, 26, 34). Available clinical data indicate that this parameter is equivalent to electrophysiological testing in assessing risk for arrhythmic events in a number of patient groups at moderate-to-high risk for cardiac events, including those with coronary artery disease, ventricular arrhythmia, myocardial infarction, cardiomyopathy, or congestive heart failure (1, 6-8, 11, 14-17, 22, 30, 36).

However, broader application of T-wave alternans analysis in the standard clinical applications of exercise treadmill testing and ambulatory electrocardiogram (ECG) has been limited. This constraint has been, in part, attributable to the intrinsic properties of spectral analytic methods that have been employed. Spectral analysis is relatively intolerant of changes in data stationarity and motion artifact and generally requires stabilizing heart rate for ~2 min (12, 29). The stationarity requirement is also problematic because major arrhythmias are often precipitated by transient physiological events, such as heightened sympathetic nerve activity, acute myocardial ischemia and reperfusion, and episodes of intense physiological or mental stress (5, 21, 23).

The goal of the present study was to develop a robust means to assess T-wave alternans that would be compatible with routine clinical monitoring and experimental laboratory studies. To optimize the extent of use, the method would need to provide interpretable results without controlling heart rate and would handle artifacts and noise. We also required that the method be dynamic, i.e., be sensitive to transient changes in cardiac electrical instability. To achieve these objectives, we developed a new technique, which we termed modified moving average analysis, formerly named "median beat" analysis (27, 38). The approach and algorithms are described together with validation studies. In addition, experimental laboratory data are provided that demonstrate the capability of the method to predict the occurrence of VF during acute coronary artery occlusion.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Modified moving average computed beat construction and analysis of T-wave alternans. The underlying assumption in our development of modified moving average analysis was that the most robust means of quantifying T-wave alternans would disclose its predictive power most fully. Accurate measurement is difficult, as alternans magnitude can range from <20 µV to several hundred microvolts on a surface ECG. The approach involves constructing modified moving average computed beats by averaging alternate ECG beats. A weighted moving average is applied to limit the contribution of any one beat. The alternans estimate for any ECG segment is then determined as the maximum difference between A and B modified moving average computed beats within the ST segment and T-wave region. These strategies allow the algorithm to discriminate surges in alternans attributable to physiological and pathophysiological triggers. The accuracy of the algorithm in quantifying T-wave alternans was verified with simulated ECGs, and its predictive accuracy was established in an experimental study.

(1a)

(1b)

(2a)

(2b)

(3a)

n1

n1

(3b)

n1

n1

View larger version (33K): [in this window] [in a new window]   Fig. 1.   Flow chart of the major components of the modified moving average (MMA) method. The left ventricular (LV) electrocardiogram (ECG) was obtained from a representative experiment in which coronary artery occlusion subsequently resulted in ventricular fibrillation (VF). ECGs are filtered to reduce high-frequency noise and to remove baseline wander. Ventricular and supraventricular premature ECG beats, as well as ECG beats with a high-noise level, are removed. The even ECG beats in the sequence are then assigned to group A and the odd ECG beats to group B. MMA computed beats of types A and B are computed continuously according to Eqs. 3a and 3b. The nth MMA computed beat is developed from the n  1th ECG beat and the n  1th MMA computed beat. The amplitude of the effect of any one ECG beat on the MMA computed beat is also limited by the bounds on A and B, where  is change, which are computed by Eqs. 3a and 3b. The alternans estimate is determined as the maximum absolute difference between A and B MMA computed beats within the ST segment and T-wave region. The output period can be adjusted as desired.

(4)

Testing with simulated ECGs. The precision of the program during dynamically changing conditions was tested by comparing known alternans input values in the simulated ECG with the output alternans readings. A simulated ECG was generated by a C++ program. The R wave, T wave, P wave, and ST segment were approximated by geometric shapes whose relative timing and amplitude were similar to that of a surface ECG. A pulse with an amplitude of 10-1,000 µV was added to alternate ECG beats of the simulated waveform and centered ~30 ms before the T-wave apex to simulate an ECG with T-wave alternans (Fig. 2). This location was chosen because the first one-half of the T wave is the period of enhanced dispersion of repolarization (20) when maximum alternans occurs (24, 26). Detection of T-wave alternans in this range of amplitudes has been sufficient for our experimental and clinical studies. We also assessed the algorithm's capacity to measure alternans during abrupt increases in alternans level (Fig. 3), alternans phase reversals (Fig. 4), changing heart rates (Fig. 5), and stray single-beat spikes (Fig. 6), as these phenomena are typical of routine clinical and experimental ECGs.

View larger version (22K): [in this window] [in a new window]   Fig. 2.   Calibration curve demonstrating a linear relationship between the alternans value estimated by MMA analysis and the alternans input signal in a simulated waveform. The alternans level measured by the MMA method is plotted against the input alternans level in a simulated waveform, which was elevated from 10 to 1,000 µV in a staircase fashion, changing once every 10 min. Regression analysis yielded a slope of 0.996, an intercept of 4.33 µV, and correlation of r2 = 0.9999. ECGs with 10-, 100-, and 1,000-µV alternans are also shown. Alternans of 10 µV is below the level that can readily be detected by the human eye, whereas 100-µV alternans is visually detectable but small. Alternans of 1,000 µV is large and is seen in only extreme cases and syndromes.

View larger version (29K): [in this window] [in a new window]   Fig. 3.   Accuracy of MMA analysis in tracking a stepwise change in alternans level. For a 50- or 100-µV step, the MMA algorithm registers 100% of the input step during the first 15-s measurement interval. For a 200-, 500-, or 1,000-µV step, the algorithm produces a result that is 87.97, 87.49, and 86.67% of the input step value during the first 15 s and 100, 100, and 99.01% of the input step value during the second 15 s, respectively. The stepwise change from 0 to 1,000 µV is tracked with 100% accuracy by the third 15-s interval. This level of accuracy has proven adequate for biological data. An even faster response time can be obtained at the expense of some noise suppression capability.

View larger version (29K): [in this window] [in a new window]   Fig. 4.   Capacity of the MMA method to handle phase reversals in the alternans pattern. Arrhythmias can sometimes trigger a phase reversal so that the alternans pattern changes from ABABAB to BABABA. An uncorrected phase reversal that occurs in the center of the data in Fourier transform analysis of alternans can mistakenly reduce the alternans estimate to zero. By contrast, the MMA method recovers completely in one or two 15-s measurement intervals. For a 50- or 100-µV alternans signal, the accuracy of the reading for the first 15-s interval after the phase reversal remains at 100%. Whereas, during the initial 15-s interval, the accuracy drops to 75% for a 200- or 500-µV alternans signal and to 73% for 1,000-µV alternans signal, it recovers during the second 15-s interval to 100% for the 200- or 500-µV alternans signal and to 98% for the 1,000-µV signal. During the third 15-s interval, all readings exhibited 100% accuracy.

View larger version (25K): [in this window] [in a new window]   Fig. 5.   Minimum change in T-wave alternans level as a result of varying heart rates (HR) from 60 to 180 beats/min (bpm). An alternans pulse of constant amplitude (100 µV) was added to every other beat of the simulated ECG at 30 ms before the T-wave apex. The T-wave (TW) apex is described by the equation TW apex = 374  1.985 · HR  0.0036 · HR2, and the T-wave end follows the equation TW end = 455  1.878 · HR  0.0026 · HR2 (39). The T-wave width and alternans pulse width are approximated as 2 · (TW end  TW apex). The width and timing of the pulse were allowed to vary with HR similar to a surface ECG. Whereas the alternans apex changes from 237.86 to 103.34 ms after the R wave as the HR varies from 60 to 180 beats/min, the measured alternans peak varied only from 96.65 to 104.61% of its true value of 100 µV.

View larger version (20K): [in this window] [in a new window]   Fig. 6.   Minimum effect on alternans readings from single ECG beats typical of motion or noise artifact. A simulated ECG beat spike of 1,000 µV produces only a 12.9-µV alternans reading, which is 1.29% of the input value. Similar results are shown for simulated ECG beat spikes of 500, 200, 100, 50, 20, or 10 µV, which produce alternans readings at 0.0014, 0.002, 0.013, 0.012, and 0% of the input value, respectively. Thus the MMA method suppresses inaccuracies in alternans values due to ECG spikes while tracking continuous alternans patterns, which typically exhibit constant or slowly varying amplitudes.

Experimental studies. The program's capacity to assess vulnerability to VF during experimental coronary artery occlusion was examined in 13 consecutive adult mongrel dogs of either sex weighing 23 ± 2 kg (Figs. 7, 8, and 9). The animals were premedicated with morphine sulfate (2 mg/kg sc) and anesthetized with -chloralose (150 mg/kg iv) with supplemental doses of -chloralose given as required. A left thoracotomy was performed at the fourth intercostal interspace. A Doppler flow probe was placed around the left anterior descending coronary artery, and a 2-0 silk snare was placed around the artery to allow occlusion of the vessel. Aortic blood pressure was measured with a Gould-Statham P50 pressure transducer. A left ventricular (LV) endocardial ECG was obtained using a 7-French USCI quadripolar catheter with an interelectrode distance of 10 mm and electrode width of 2 mm. The tip of the catheter was positioned in the apex of the LV through a carotid artery. A pigtail pressure catheter was positioned to monitor LV blood pressure. Arterial blood pH, the partial pressure of CO₂, and the partial pressure of O₂ were monitored with an Instrumentation Laboratory 1304 blood-gas analyzer. Values were maintained within physiological ranges by adjusting the ventilation of the Harvard respirator.

View larger version (15K): [in this window] [in a new window]   Fig. 7.   T-wave alternans in LV ECG determined by MMA analysis in a representative experiment in which VF did not occur during 8 min of occlusion and abrupt reperfusion of the left anterior descending coronary artery, with atrial pacing at 150 beats/min. Characteristically, alternans magnitude increased during occlusion, and a larger bidirectional alternating T-wave pattern appeared during reperfusion. The maximum coronary artery occlusion-induced alternans level in this animal was 0.383 mV, well below the receiver-operator characteristic cut point of 0.75 mV and the 1.84 ± 0.29 mV level observed in experiments in which occlusion triggered VF.

View larger version (25K): [in this window] [in a new window]   Fig. 8.   Pronounced T-wave alternans in LV endocardial tracings at ~3.5 min after the start of occlusion in a representative animal in which VF occurred (top). The amplitude of the LV blood pressure (LVBP) waveform (middle) decreased by ~10% during occlusion but did not alternate, indicating that mechanical alternans did not contribute to the appearance of T-wave alternans. Alternans was minimum in the ECG obtained at the same time point during occlusion from an animal in which VF did not occur (bottom). T-wave alternans was not visible at preocclusion baseline in any animal. To exclude a contribution to alternans levels by HR, all animals were paced at 150 beats/min.

View larger version (21K): [in this window] [in a new window]   Fig. 9.   Ischemia-induced T-wave alternans (TWA) in LV endocardial lead was significantly higher in the 6 animals in which VF occurred than in the 7 animals without VF during coronary artery occlusion (1.84 ± 0.29 vs. 0.38 ± 0.09 mV; P < 0.001) but not at baseline (0.31 ± 0.08 vs. 0.20 ± 0.02 mV; not significant). Ischemia induced an increase in alternans in the animals in which VF occurred (1.84 ± 0.29 vs. 0.31 ± 0.08 mV at baseline) but not in the animals in which VF did not occur (0.38 ± 0.09 vs. 0.20 ± 0.02 mV at baseline; not significant). Alternans was measured at 3.6 ± 0.2 min, and VF occurred 4.37 ± 0.15 min after the start of left anterior descending coronary artery occlusion. To provide data within the working range of the algorithm, the ECG signal was scaled down by a factor of 10 to the size of alternans in a surface ECG. The actual intracavitary alternans values were determined by restoring the factor of 10. Ischemia-induced alternans in LV endocardial lead was >0.75 mV in all 6 animals in which VF occurred, whereas, in 7 animals in which VF did not occur, alternans remained <0.75 mV. The cut point was determined from the receiver-operator characteristic curve to be 0.75 mV, where both specificity and sensitivity were 100%.

Preparation of experimental laboratory tracings. Recording and analysis of data were performed with commercial equipment (GE Medical Systems Information Technologies, Milwaukee, WI). Experimental laboratory ECG data recorded with standard electrodes were low-pass filtered at 50 Hz before sampling at 500 Hz per channel and stored on rewritable optical disks by Streamer software. These ECG data were down-sampled to 125 Hz for loading on the MARS workstation for analysis. The MARS workstation is designed to handle surface ECGs with a range of 5 to 5 mV with least significant bit resolution of 2.44 µV (12-bit analog/digital). Because the R-wave amplitude of the LV ECG is larger than that of the surface ECG, we scaled down the LV ECG by a factor of 10 for analysis. Other scaling approaches can be employed. Ectopic beats, ventricular arrhythmias, or artifacts identified by the MARS workstation were verified by a trained operator and removed from analysis.

Statistical methods. The statistical tests were carried out with a STATA statistical package (Santa Monica, CA). Results are expressed as means ± SE. A linear regression analysis was performed (Fig. 2). An ANOVA was performed to calculate the differences between the two groups of experimental laboratory data (Figs. 8 and 9). Specificity and sensitivity were calculated with standard formulas (Fig. 9). The cut point was determined from the receiver-operator characteristic (ROC) curve, where both specificity and sensitivity were 100%, to be 0.75 mV.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Simulation studies. Validation testing with the simulated alternans signal, added to increase alternans in a stepwise fashion from 10 to 1,000 µV, determined that this relatively simple algorithm measured T-wave alternans as the maximum difference in T-wave amplitude of successive ECG beats with a high degree of accuracy across this range of values. We observed a correlation coefficient of 0.9999, indicating a significant level of precision (Fig. 2). The algorithm performed accurately from 1,000 to 10 µV, which has been an adequate lower limit in our studies. Alternans levels <10 µV were not tested because the current least significant bit resolution of the MARS workstation is 2.44 µV. However, lower levels of T-wave alternans could be detected with equipment using analog-to-digital converters with higher resolution. The program assessed abrupt increases in alternans level of 50-1,000 µV, demonstrating its capacity to track changes in alternans level, such as those that occur in response to physiological triggers (Fig. 3). The algorithm also accurately handled sudden phase reversals or resetting that can result from arrhythmias (Fig. 4). This capability was demonstrated by adding an abrupt phase reversal in the alternans pattern from ABABAB to BABABA. The magnitude of alternans varied <5% from its input value as heart rate was raised from 60 to 180 beats/min, and the width and timing of the simulated alternans pulse were allowed to change realistically with heart rate (Fig. 5). The method discriminated between an isolated ECG beat spike, such as that which is commonly due to motion artifact, and an alternans signal; this capacity was tested by adding an alternans pulse to a single ECG beat in a stream of otherwise constant-amplitude beats (Fig. 6).

Experimental laboratory study. Modified moving average analysis revealed a progressive increase in T-wave alternans magnitude at 3-4 min and a subsequent abrupt rise in alternans level on release reperfusion at 8 min after coronary artery occlusion (Fig. 7). This pattern coincides with previous reports of the time course of cardiac vulnerability in response to ischemia and reperfusion, as tracked by programmed electrical stimulation (20) and spontaneous occurrence of ventricular tachyarrhythmias (20, 26). Modified moving average algorithm demonstrated that T-wave alternans was capable of differentiating the animals that subsequently experienced VF from those that did not fibrillate during coronary artery occlusion (Figs. 8 and 9). Alternans levels were measured at 3.6 ± 0.2 min, and fibrillation ensued at 4.37 ± 0.15 min after the start of the occlusion in susceptible animals. This predictive ability is independent of heart rate, which was held constant at 150 beats/min by atrial pacing. Analysis of the LV blood pressure waveform revealed no evidence of mechanical alternans. The predictive capability was exceptionally high, as the sensitivity and specificity for indicating susceptibility to ischemia-induced VF were both 100% at the cut point of 0.75 mV (Fig. 9).

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

The main goal of the present study was to develop a robust method for T-wave alternans analysis to optimize its predictive accuracy for life-threatening arrhythmias. The critical performance characteristics required were reliability of measurement during abrupt changes in alternans magnitude and phase reversals and insensitivity to changes in data stationarity, including heart rate fluctuations and ECG beat spikes. These ECG phenomena are commonly encountered during acute myocardial ischemia and reperfusion, as well as altered autonomic nervous system activity. Modified moving average analysis is a nonspectral method that builds on the powerful noise-reduction principle of signal averaging by estimating moving average computed beats (Fig. 1). Experimental laboratory data are provided that document the high predictive capacity of the method in measuring T-wave alternans to assess risk for VF during acute coronary artery occlusion.

ECG simulation studies. The calibration curve (Fig. 2) demonstrated a strong linear relationship (r² = 0.9999) between the T-wave alternans value estimated by modified moving average analysis and the input signal of a simulated ECG waveform with increasing alternans magnitudes. The algorithms also performed well in tracking stepwise changes in alternans level (Fig. 3) and handling phase reversals in alternans patterns (Fig. 4). Importantly, we observed minimum effects on alternans measurements from increases in heart rate (Fig. 5) or from single-beat spikes (Fig. 6), which are typical of motion or noise artifact and are commonly encountered during exercise (27) and ambulatory monitoring (28, 38). The approach accurately measures bidirectional alternans, with the T-wave inscribed alternately above and below the isoelectric line, which discloses heightened vulnerability such as may occur on reperfusion of ischemic myocardium (24, 26) or exercise and emotional stress in patients with the Long QT Syndrome (4, 33).

Electrophysiological considerations. In the present study, modified moving average analysis of T-wave alternans was demonstrated to be capable of tracking vulnerability to VF during myocardial ischemia with a high degree of sensitivity and specificity (Figs. 7-9). This desirable result probably reflects both the fundamental electrophysiological link between T-wave alternans and vulnerability to VF (24, 26, 36, 37) and the capacity of the present method to assess the level of T-wave alternans accurately. Although the precise mechanisms of ischemia-induced T-wave alternans remain to be determined, the phenomenon appears to reflect the degree of heterogeneity of repolarization, an electrophysiological property that has been extensively implicated in the genesis of VF (9, 36, 37). The close coupling between T-wave alternans and vulnerability to VF is also supported by the characteristic crescendo in alternans magnitude observed during transition from normal rhythm to VF in both the present (Fig. 8) and previous studies (19, 24, 26). In animals that did not fibrillate during coronary artery occlusion, a low level of alternans was registered (Figs. 7-9), but its magnitude remained below the level found in their vulnerable counterparts. The lower level of alternans in the animals that did not fibrillate may be attributable to a number of factors, including differences in the extent of preformed collateral channels, baroreceptor sensitivity, and the degree of sympathetic nerve activation in response to myocardial ischemia (32). Accurate detection of alternans levels was probably optimized by the use of electrodes positioned in the LV near the site of ischemia, because ischemia-induced T-wave alternans is a regionally specific phenomenon (18, 26, 35). In humans and animals, even during severe myocardial ischemia, local electrograms from outside the zone of ischemia reveal little or no T-wave alternans. It remains unknown, however, whether the ideal sensitivity and specificity observed in the present study, with data obtained with LV electrograms, will be matched using body surface electrodes, which yield a lower signal-to-noise ratio.

Performance of modified moving average analysis compared with complex demodulation. In previous studies using complex demodulation, we demonstrated that T-wave alternans magnitude correlates with susceptibility to VF under diverse physiological and pharmacological interventions (19, 24-26). The desirable ROCs obtained with the modified moving average algorithm in analyzing experimental laboratory ECGs monitored with identical ECG recording systems indicate that its signal processing features are superior to those of complex demodulation (26) (Fig. 9). The ROC curve for predicting VF using surface ECGs, which yield a less favorable signal-to-noise ratio than LV ECGs, remains to be determined. However, based on our experience with ambulatory ECGs (38) and treadmill testing (27), modified moving average analysis is likely to be preferable to complex demodulation, which can be disrupted by motion and noise artifact (unpublished observations).

Preliminary clinical studies. During routine symptom-limited exercise treadmill testing, modified moving average analysis of T-wave alternans differentiated patients with stable coronary artery disease from age-matched normal volunteers (27). At exercise treadmill testing heart rates of 120 beats/min, alternans was higher in patients than in normal subjects (55.7 ± 5.5 vs. 37.4 ± 4.7 µV; P < 0.05; n = 7 for both groups), although there was no difference in alternans levels between the groups at preexercise baseline (11.9 ± 1.4 µV for patients vs. 15.5 ± 2.3 µV for normal subjects; not significant). Interpretable values were obtained in all cases during both rest and activity without specialized ECG electrodes or protocols to maintain heart rate constant. Modified moving average analysis of T-wave alternans is also promising in terms of predicting a significant increase in the risk of arrhythmic death (38). These results were obtained from routine 24-h ambulatory ECG tapes from a moderate-risk group of 1,284 postmyocardial infarction patients enrolled in the Autonomic Tone and Reflexes After Myocardial Infarction multicenter study. Using a nested-case control study design, we analyzed, in a blinded fashion, 14 cases and 25 controls matched for gender, age, site of myocardial infarction, LV ejection fraction, and thrombolysis. A statistically significant increase in risk of arrhythmic death was predicted by T-wave alternans level (P < 0.05). Thus results of the predictive power T-wave alternans obtained using modified moving average analysis in standard exercise treadmill testing and ambulatory monitoring without specialized electrodes are encouraging, although preliminary.

Conclusions and implications. Laboratory testing of simulated ECG signals and experimental laboratory data indicate that modified moving average analysis possesses suitable performance characteristics for evaluating the impact of physiological and pathophysiological factors on T-wave alternans. This method also demonstrates the high predictive accuracy of T-wave alternans for estimating risk for life-threatening ventricular tachyarrhythmias. In addition, because effective antiarrhythmic agents suppress T-wave alternans (8, 16, 19, 25, 31) and the phenomenon has been reported to precede drug-induced torsade de pointes (3, 10, 13), there is a potential application in drug testing.