Cardiovascular Engineering

, Volume 9, Issue 1, pp 18–26 | Cite as

Spectral Energy of ECG Morphologic Differences to Predict Death

  • Zeeshan Syed
  • Phil Sung
  • Benjamin M. Scirica
  • David A. Morrow
  • Collin M. Stultz
  • John V. Guttag
Original Paper

Abstract

Unstable conduction system bifurcations following ischemia and infarction are associated with variations in the electrocardiographic activity spanning the heart beat. In this paper, we investigate a spectral energy measure of morphologic differences (SE-MD) that quantifies aspects of these changes. Our measure uses a dynamic time-warping approach to compute the time-aligned morphology differences between pairs of successive sinus beats in an electrocardiographic signal. While comparing beats, the entire heart beat signal is analyzed in order to capture changes affecting both depolarization and repolarization. We show that variations in electrocardiographic activity associated with death can be distinguished by their spectral characteristics. We developed the SE-MD metric on holter data from 764 patients from the TIMI DISPERSE2 dataset and tested it on 600 patients from the TIMI MERLIN dataset. In the test population, high SE-MD was strongly associated with death over a 90 day period following non-ST-elevation acute coronary syndrome (HR 10.45, p < 0.001) and showed significant discriminative ability (c-statistic 0.85). In comparison with heart rate variability and deceleration capacity, SE-MD was also the most significant predictor of death in the study population. Furthermore, SE-MD had low correlation with these other measures, suggesting that complementary use of the risk variables may allow for more complete assessment of cardiac health.

Keywords

Electrocardiogram (ECG) Risk stratification Acute coronary syndromes Heart rate variability Deceleration capacity Morphologic differences 

Introduction

In a stationary and homogenous myocardial conducting system, the activated pathways through excitable cells are usually similar for consecutive beats. However, in the presence of ischemia, the conducting system has multiple irregular islands of severely depressed and unexcitable myocardium (El-Sherif et al. 1977) that leads to discontinuous electrophysiological characteristics (Josephson and Wit 1984). The presence of several possible adjacent pathways that can invade the nonfunctioning area leads to variations in the spatial direction of the invading vector (Ben-Haim et al. 1991). Measured electrical activity in this phase can only be described in probabilistic terms because of beat-to-beat activation and repolarization variability, stemming from subtle unstable conduction bifurcations. Furthermore, propagation of a beat may be dependent on the route of propagation of the previous beat. The overall effect of such minor conduction inhomogeneities is not well understood, but it is possible that they correlate with myocardial electrical instability and have potentially predictive value for ventricular arrhythmias (Ben-Haim et al. 1991) or other adverse events.

In this paper, we propose and evaluate a new method to risk stratify patients by measuring seemingly random morphologic differences in electrocardiographic (ECG) signals. Our method uses dynamic time-warping to compute the time-aligned morphology changes between consecutive sinus beats. While comparing beats, the entire heart beat signal is analyzed and changes affecting both depolarization and repolarization are quantified. This approach reduces the original electrocardiographic signal to a time series of morphologic differences (MD). We describe how, when a large enough sequence of beats is analyzed, increased spectral energy in the MD time series (SE-MD) within a characteristic frequency range may have value in predicting patients at increased future risk of cardiovascular death.

We developed our SE-MD metric on a training set of 764 patients from the TIMI DISPERSE2 study, who were followed up for the endpoint of death for a 90 day period following non-ST-elevation acute coronary syndromes (NSTEACS). We then tested SE-MD on 600 patients from the TIMI MERLIN dataset. For both these datasets, SE-MD identified patients at increased risk of a future adverse outcome. SE-MD was also more strongly associated with death in the test dataset than either heart rate variability (HRV) or deceleration capacity (DC). Furthermore, SE-MD had low correlation with these other measures, suggesting that complementary use of these risk variables may allow for more complete assessment of cardiac health.

The rest of this paper is organized as follows. “Morphologic Differences” details the process of measuring time-aligned morphology differences between pairs of consecutive heart beats. “Spectral Energy of Morphologic Differences” describes the approach of identifying characteristic frequencies in the MD time series that are associated with future cardiovascular death. “Evaluation” proposes a study to evaluate the prognostic information provided by SE-MD, HRV and DC. “Results presents the results of this study. “Summary and Discussion” concludes the paper with a summary and discussion.

Morphologic Differences

This section describes the different stages involved in calculating the MD time series.

ECG Signal Preprocessing

The process of analyzing ECG morphology is more sensitive to noise than techniques focusing exclusively on the heart rate. This is because the heart rate can often be estimated robustly, even in the presence of significant amounts of noise, by searching for high amplitude R-waves in the signal. In contrast, characterizing the morphology requires using information even from those parts of the ECG that are low amplitude and where small amounts of noise can significantly affect the signal-to-noise ratio. To minimize this effect, we employ different techniques for noise removal and automated signal rejection.

Noise removal is carried out in three steps. Baseline wander is first removed by subtracting an estimate of the wander obtained by median filtering the original ECG signal as described in (DeChazal et al. 2004). The ECG signal is then filtered using wavelet denoising with a soft-threshold (Donoho 2005). Finally, sensitivity to calibration errors is decreased by normalizing the entire ECG signal by the mean R-wave amplitude.

While the noise removal steps help remove artifacts commonly encountered in long-term electrocardiographic records, the signal rejection process is designed to remove segments of the ECG signal where the signal-to-noise ratio is sufficiently low that meaningful analysis of the morphology is challenging even after noise removal. Such regions are typically dominated by artifacts unrelated to cardiac activity but that have similar spectral characteristics to the ECG signal, e.g., segments recorded during periods when there was substantial muscle artifact.

The process of signal rejection proceeds in two stages. Parts of the ECG signal with a low signal quality index (Li et al. 2008) are identified by combining four analysis methods: disagreement between multiple beat detection algorithms on a single ECG lead, disagreement between the same beat detection algorithm on different ECG leads, the kurtosis of a segment of ECG, and the ratio of power in the spectral distribution of a given ECG segment between 5–14 and 5–50 Hz. In our work, we use the Physionet SQI package implementation (Li et al. 2008) to automatically remove parts of the ECG signal with a low signal quality index from further analysis. The remaining data is divided into half hour windows, and the standard deviation of the R-waves during each half hour window is calculated. We discard any window with a standard deviation greater than 0.2887. Given the earlier normalization of the ECG signal, under a conservative model that allows the R-wave amplitude to uniformly vary between 0.5 and 1.5 every beat (i.e., up to 50% of its mean amplitude), we expect the standard deviation of the R-wave amplitudes to be less than 0.2887 for any half hour window. This heuristic identifies windows that are likely corrupted by significant non-physiological additive noise, and where the morphology of the ECG cannot be meaningfully analyzed.

ECG Segmentation and Removal of Ectopy

To segment the ECG signal into beats, we use two open-source QRS detection algorithms with different noise sensitivities. The first of these makes use of digital filtering and integration (Hamilton and Tompkins 1986) and has been shown to achieve a sensitivity of 99.69%, while the second is based on a length transform after filtering (Zong et al. 2003) and has a sensitivity of 99.65%. Both techniques have a positive predictivity of 99.77%. QRS complexes were marked only at locations where these algorithms agreed. These algorithms are used as part of the Physionet SQI package implementation described earlier.

In order to study sinus conduction, prior to further analysis, ectopic parts of the signal were also removed in a fully automated manner using the beat classification algorithm of (Hamilton 2002) present in the Physionet SQI package. The beat classification algorithm characterizes each beat by a number of features such as width, amplitude and RR interval, and then compares it to previously detected beat types to assign it a label.

We removed each ectopic beat, as well as the beats occurring immediately before and after it. To address the splicing introduced by this step, we also made changes to the subsequent stages of measuring SE-MD. We restricted the process of measuring pairwise energy differences (“Morphologic Distance (MD) Time Series”) to pairs of beats that occurred consecutively, i.e., did not span a gap corresponding to spliced out beats. We also used the Lomb-Scargle periodogram to measure spectral energy (“Spectral Energy of Morphologic Differences”), since this method has been shown to be well-suited for irregularly sampled signals (Clifford and Tarassenko 2005).

Morphologic Distance (MD) Time Series

For every pair of consecutively occurring beats for a patient, we develop a metric that quantifies how the ECG morphology between these beats differs. The simplest way to calculate this energy difference is to subtract the samples of one beat from another. However, if samples are compared based strictly on their distance from the start of the beat, this process may end up computing the differences between samples associated with different waves or intervals. For example, consider the two heart beats depicted in Fig. 1. In the drawing on the left, samples are aligned based on their distance from the onset of the P-wave. One consequence of this approach is that samples that are part of the T-wave of the top beat are compared with samples not associated with the T-wave of the bottom beat. A measure computed this way may not reflect differences in the shapes of the T-waves of adjacent beats.
Fig. 1

Alignment of beats by dynamic time-warping. Vertical lines connect corresponding samples from the beat colored upper to the beat colored lower. If samples are compared strictly on their distance from the start of the beat, this process may end up computing the difference between unrelated parts of the two beats. Consider, for example, the two heart beats shown. In the drawing on the left, samples are aligned based on their distance from the onset of the P-wave. One consequence of this is that samples that are part of the T-wave of the top beat are compared with samples not associated with the T-wave of the bottom beat. A measure computed this way will not reflect differences in the shapes of the T-waves of adjacent beats. As depicted in the drawing on the right side, this can require aligning a single sample in one beat with multiple samples in another beat. The dynamic time-warping algorithm produces the optimal alignment of two sequences of possibly different lengths, where optimality is defined as minimizing the overall distortion due to differences in amplitude and timing of ECG waves

We use a variant of dynamic time-warping (DTW) (Rabiner 1978) to align samples that correspond to the same underlying physiological activity. As depicted in the drawing on the right side of Fig. 1, this can require aligning a single sample in one beat with multiple samples in another beat. The algorithm uses dynamic programing to search for an alignment that minimizes the overall distortion. Distortion is measured using the method described in (Syed et al. 2007), which captures differences in both amplitude and timing of ECG waves.

More precisely, given two beats, x1 and x2, of length l1 and l2 respectively, DTW produces the optimal alignment of the two sequences by first constructing an l1-by-l2 distance matrix d. Each entry (i, j) in this matrix d represents the square of the difference between samples x1[i] and x2[j]. A particular alignment then corresponds to a path, φ through the distance matrix of the form:
$$ \varphi (k) = (\varphi_{1} (k),\varphi_{2} (k)),\quad 1 \le k \le K $$
(1)
where φ1 and φ2 represent row and column indices into the distance matrix, and K is the alignment length. Any feasible path must obey the endpoint constraints:
$$ \varphi_{1} (1) = \varphi_{2} (1) = 1 $$
(2)
$$ \varphi_{1} (K) = l_{1} $$
(3a)
$$ \varphi_{2} (K) = l_{2} $$
(3b)
as well as the continuity and monotonicity conditions:
$$ \varphi_{1} (k + 1) \le \varphi_{1} (k) + 1 $$
(4a)
$$ \varphi_{2} (k + 1) \le \varphi_{2} (k) + 1 $$
(4b)
$$ \varphi_{1} (k + 1) \ge \varphi_{1} (k) $$
(5a)
$$ \varphi_{2} (k + 1) \ge \varphi_{2} (k) $$
(5b)
The optimal alignment produced by DTW minimizes the overall cost:
$$ C(x_{1} ,x_{2} ) = \mathop { \min }\limits_{\varphi } C_{\varphi } (x_{1} ,x_{2} ) $$
(6)
where Cφ is the total cost of the alignment path φ and is defined as:
$$ C_{\varphi } (x_{1} ,x_{2} ) = \sum\limits_{k = 1}^{K} d (x_{1} [\varphi_{1} (k)],x_{2} [\varphi_{2} (k)]) $$
(7)
The search for the optimal path is carried out in an efficient manner using a dynamic programing algorithm derived from the following recurrence for the cumulative path distance, γ(i, j), and the distance matrix d:
$$ \gamma (i,j) = d(i,j) + { \min }\left\{ {\begin{array}{*{20}c} {\gamma (i - 1,j - 1)} \hfill \\ {\gamma (i - 1,j)} \hfill \\ {\gamma (i,j - 1)} \hfill \\ \end{array} } \right\} $$
(8)

The final energy difference between the two beats x1 and x2, is given by the cost of their optimal alignment, which depends on the amplitude differences between the two signals and the length, K, of the alignment (which increases if the two signals differ in their timing characteristics). In a typical formulation of DTW, this difference is divided by K to remove the dependence of the cost on the length of the original observations. A problem with applying this correction in our context is that some paths are long not because the segments to be aligned are long, but rather because the observations are time-warped differently. In these cases, dividing by K is inappropriate since a difference in the length of a beats (or of parts of beats) often provides diagnostic information that is complementary to the information provided by the morphology. Consequently, in our algorithm we omit the division by K.

A further modification to traditional DTW in our work is that we restrict the local range of the alignment path in the vicinity of a point to prevent biologically implausible alignments of large parts of one beat with small parts of another. For example, for an entry (i, j) in the distance matrix d, we only allow valid paths passing through (i−1, j−1), (i−1, j−2), (i−2, j−1), (i−1, j−3) and (i−3, j−1). This is an adaptation of the Type III and Type IV local continuity constraints proposed in (Myers et al. 1980) and ensures that there are no long horizontal or vertical edges along the optimal path through the distance matrix, corresponding to a large number of different samples in one beat being aligned with a single sample in the other. This leads to the following recurrence relation (which is also shown graphically in Fig. 2):
$$ \gamma (i,j) = d(i,j) + { \min }\left\{ {\begin{array}{*{20}c} {\gamma (i - 1,j - 1)} \hfill \\ {d(i - 1,j) + \gamma (i - 2,j - 1)} \hfill \\ {d(i - 1,j) + d(i - 2,j) + \gamma (i - 3,j - 1)} \hfill \\ {d(i,j - 1) + \gamma (i - 1,j - 2)} \hfill \\ {d(i,j - 1) + d(i,j - 2) + \gamma (i - 1,j - 3)} \hfill \\ \end{array} } \right\} $$
(9)
Fig. 2

Traditional and modified recurrence relation of dynamic time-warping

The process described above transforms the original ECG signal from a sequence of beats to a sequence of energy differences. We call the resulting time series the morphologic distance (MD) time series for the patient. This new signal, comprising pair-wise, time-aligned energy differences between beats, is then smoothed using a median filter of length 8. The median filtering process addresses noisy and ectopic heart beats that may have passed through the earlier preprocessing stage and lead to high morphologic distances. The smoothing process is geared towards ensuring that high values in the MD time series correspond to locally persistent morphology changes, i.e., sustained differences in beat-to-beat morphology.

Spectral Energy of Morphologic Differences

We estimate the power spectral density of the MD time series using the Lomb-Scargle periodogram (Lomb 1976), which is well-suited to measure the spectral content of an irregularly sampled signal by taking into account both the signal value and the time of each sample. The Lomb-Scargle periodgram provides a natural way to exclude noisy samples from the computation, unlike other spectral estimation techniques that require interpolation methods to deal with missing data.

For a time series where the value m[n] is sampled at time t[n], the Lomb-Scargle periodogram giving the energy at frequency ω is defined as:
$$ P(\omega ) = \frac{1}{{2\sigma^{2} }}\left\{ {\frac{{\sum\limits_{n} {[(m[n] - \mu )\cos \,\omega (t[n] - \tau )]^{2} } }}{{\sum\limits_{n} {\cos^{2} \omega (t[n] - \tau )} }} + \frac{{\sum\limits_{n} {[(m[n] - \mu )\sin \,\omega (t[n] - \tau )]^{2} } }}{{\sum\limits_{n} {\sin^{2} \omega (t[n] - \tau )} }}} \right\} $$
(10)
where μ and σ are the mean and variance of the m[n], and τ is defined as:
$$ { \tan }(2\omega \tau ) = \frac{{\sum\nolimits_{n} {\sin (2\omega t[n])} }}{{\sum\nolimits_{n} {\cos (2\omega t[n])} }} $$
(11)
To distinguish between pathological variations in morphology and those resulting from noise or non-pathological physical effects, we investigated a frequency range within the power spectral density of the MD time series that had maximal prognostic information. This was done by evaluating all possible frequency bands within 0.1–0.6 Hz with a granularity of 0.01 Hz in a training set of data. We used ECG recordings from 764 patients in the TIMI DISPERSE2 study (Cannon et al. 2007) for this purpose. Patients in the DISPERSE2 study had 24 h holter ECG data recorded at 128 Hz within 48 h of admission due to NSTEACS. 15 deaths were observed in this population over a follow-up period of 90 days. For each possible choice of prognostic frequencies, we computed energy in the MD power spectrum for all patients in the study population. These ranges were evaluated for their ability to discriminate between those patients who died and those who did not by measuring the corresponding c-statistic (i.e., the area under the receiver operating characteristic curve) (Ohman et al. 2000). As shown in Fig. 3, the frequency range of 0.30–0.55 Hz yielded the maximum prognostic information in the study population and led to a training set c-statistic of 0.77. Based on this experiment, we define SE-MD as the energy in the MD time series power spectral density between 0.30 and 0.55 Hz.
Fig. 3

Heatmap showing c-statistic as a function of low- and high-frequency cutoffs. A maximum c-statistic value of 0.77 is obtained for the frequency band 0.30–0.55 Hz, i.e., the spectral energy (SE-MD) band

Evaluation

We tested the ability of SE-MD to discriminate between low and high risk patients in a study on 600 patients randomly selected from the placebo population for the TIMI MERLIN (Morrow et al. 2007) trial. Each of these patients had 24 h of holter ECG recorded at 128 Hz within 48 h of admission due to NSTEACS. There were 12 cardiovascular deaths in this cohort during a follow-up period of 90 days.

For comparison, we also studied the discriminative ability of two other commonly used electrocardiographic risk variables: heart rate variability (HRV) (Malik 1996) and deceleration capacity (DC) (Bauer et al. 2006). For HRV, we considered the SDNN, SDANN, ASDNN, RMSSD, HRVI, pNN50 and LF/HF metrics that were proposed by the Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology (Malik 1996), but report only data from the best performing (i.e., LF/HF) metric. To measure DC, we used the PRSA implementation available at: http://www.psra.eu (Bauer et al. 2006).

To evaluate the risk variables, we carried out two separate analyses. We calculated the c-statistic for the predicted survival of patients over a 90 day period following NSTEACS. We then dichotomized the risk variables using the values reported in the literature for HRV (i.e., 1.2) (Malik 1996) and DC (2.5) (Bauer et al. 2006), and the highest quartile value for SE-MD in the DISPERSE2 dataset (52.5). For the dichotomized risk variables, we subsequently estimated hazard ratios using Cox proportional hazards regression models (Cox and Oakes 1984).

In addition to using values reported in the literature as cutoffs for dichotomization, we also estimated hazard ratios for high risk groups comprising patients in the lowest quartiles for HRV and DC (where low values correspond to high risk) and patients in the highest quartile for SE-MD (where high values correspond to high risk). This approach helped provide a comparison of the different methods on equal sized populations. We further investigated using a high risk group for each risk variable that was chosen to match the percentage of deaths observed in the TIMI DISPERSE2 dataset over the 90 day follow-up period (i.e., patients in the lowest 2.5% HRV and DC, and in the highest 2.5% SE-MD). These alternate dichotomization cutoffs only affect the hazard ratios obtained. The measurement of the c-statistic does not require dichotomization, and calculation of this value is not repeated for changes in cutoff.

The decision to calculate both c-statistics and hazard ratios was based on the complementary information provided by these statistical techniques. The c-statistics are calculated by measuring the area under the ROC curve and indicate the discriminative ability of risk variables. This data does not factor in the survival time for patients who die. Furthermore, it ignores censoring, i.e., patients dropping out before the study is complete. Conversely, the hazard ratios for univariate association are derived using a Cox proportional regression model, and relate the survival times of patients in high and low risk groups, while accounting for censoring. Intuitively, the c-statistics can be considered as a measure of discriminative ability of a risk variable, while hazard ratios report the survival characteristics of patients in high and low risk populations based on a dichotomized classification of patients using the risk variable.

Finally, in contrast to the TIMI DISPERSE2 dataset used for training, where 90 day follow-up data was available, patients in the MERLIN study were followed up for a median duration of 348 days. We therefore also investigated risk stratification of patients using the risk variables over a year following NSTEACS. During this period 22 cardiovascular deaths occurred.

Results

Univariate and Multivariate Analysis with Dichotomization Cutoffs from Literature

The results of univariate analysis for the risk variables are shown in Tables 1 and 2. On dichotomization 193 (32.1%) patients were placed into the risk group for HRV, 55 (9.2%) for DC and 135 (22.5%) for SE-MD. HRV identified 6 of the 12 cardiovascular deaths that took place during the 90 day follow-up period in its high risk group. The corresponding number for DC was 5, while for SE-MD it was 9.
Table 1

Discriminative ability of SE-MD, HRV and DC measured using the c-statistic to predict cardiovascular death over 90 days following NSTEACS (n = 600)

Parameter

c-Statistic

SE-MD

0.85

HRV

0.64

DC

0.75

Table 2

Univariate and multivariate association between risk variables and cardiovascular death over 90 days following NSTEACS (n = 600)

Parameter

Univariate hazard ratio

95% confidence interval

p value

Multivariate hazard ratio

95% confidence interval

p value

SE-MD

10.45

2.83–38.59

<0.001

9.05

2.05–39.97

0.004

HRV

2.12

0.68–6.58

0.192

0.54

0.15–2.03

0.364

DC

7.37

2.34–23.23

<0.001

3.27

0.87–12.33

0.080

Of the evaluated measures, SE-MD was strongly associated with death over a 90 day period following NSTEACS and showed the highest hazard ratio (HR 10.45, p < 0.001). DC was also associated with death in this population. The c-statistic for SE-MD was the highest of the three measures studied, and exceeded the threshold of 0.8 associated with genuine clinical utility (Ohman et al. 2000).

Kaplan–Meier mortality curves for patients in the high and low risk populations determined by SE-MD are shown in Fig. 4. Patients in the highest quartile of SE-MD during the first 24 h after study entry were at significantly elevated risk of death over the subsequent 90 days. This effect was consistent over the entire 90 day period.
Fig. 4

Kaplan–Meier cardiovascular mortality curves comparing patients in the highest quartile of SE-MD (upper) with the remaining patients in the cohort (lower)

The results of multivariate analysis including all three electrocardiographic measures are presented in Table 2. SE-MD was the only electrocardiographic risk variable independently associated with death during follow-up in this population (HR 9.05, p = 0.004). The association between DC and the endpoint of cardiovascular death marginally exceeded the threshold of significance.

We also measured the correlation between the different electrocardiographic variables. The results of this analysis are reported in Table 3. All three ECG variables showed low correlation with each other.
Table 3

Correlation coefficients between SE-MD, HRV and DC

Parameter

SE-MD

HRV

DC

SE-MD

1.00

−0.31

−0.32

HRV

 

1.00

0.24

DC

  

1.00

Analysis with Alternate Dichotomization Cutoffs

The results of univariate and multivariate analyses when high risk groups were chosen as the lowest quartiles of HRV and DC, and the highest quartile of SE-MD, are reported in Table 4. In this case, all high risk groups consisted of 150 patients. HRV identified 6 of the 12 cardiovascular deaths that took place during the 90 day follow-up in its high risk group. The corresponding number for DC was 7, while for SE-MD it was 9. As was the case for the earlier results obtained using dichotomization cutoffs reported in the literature, SE-MD showed the highest hazard ratio on univariate analysis (HR 9.14, p < 0.001) and was the only electrocardiographic risk variable associated with cardiovascular death in this population on multivariate analysis (HR 6.61, p = 0.01). While the results of HRV on univariate analysis improved over those obtained using cutoffs from the literature, for DC the trend was reversed.
Table 4

Univariate and multivariate association between risk variables and cardiovascular death over 90 days following NSTEACS using high risk quartiles for dichotomization (n = 600)

Parameter

Univariate hazard ratio

95% confidence interval

p value

Multivariate hazard ratio

95% confidence interval

p value

SE-MD

9.14

2.48–33.77

<0.001

6.61

1.53–28.58

0.011

HRV

3.08

0.99–9.54

0.052

1.15

0.33–4.06

0.823

DC

4.38

1.39–13.80

0.012

1.88

0.51–6.88

0.340

We also considered the extent to which the same patients were placed in the high risk groups by the three risk variables. This was done by studying the intersection of these equally sized high risk populations (Table 5). About half of the high risk patients were common to any pair of risk variables, while only a third were common to all three. There was more agreement if one considers only the patients who died during the study period. Five patients who died were identified as high risk by all three methods. DC identified one patient that neither of the other methods classified as high risk. SE-MD identified three patients not identified by HRV and four patients not identified by DC.
Table 5

Patients common to high risk groups for all three risk variables using high risk quartiles for dichotomization (n = 600)

Parameter

Patients

Cardiovascular deaths

SE-MD

150

9

HRV

150

6

DC

150

7

SE-MD ∩ HRV

79

6

SE-MD ∩ DC

77

5

HRV ∩ DC

77

5

SE-MD ∩ HRV ∩ DC

55

5

Table 6 reports the results of univariate and multivariate analyses when high risk population sizes were chosen to match the observed rate of death over 90 days in the TIMI DISPERSE2 dataset. The results obtained for these analyses parallel our earlier findings.
Table 6

Univariate and multivariate association between risk variables and cardiovascular death over 90 days following NSTEACS using high risk group sizes matching the rate of cardiovascular death in the TIMI DISPERSE2 dataset (n = 600)

Parameter

Univariate hazard ratio

95% confidence interval

p value

Multivariate hazard ratio

95% confidence interval

p value

SE-MD

8.10

1.77–36.97

0.007

6.93

1.23–39.10

0.028

HRV

0.00

0.00

DC

3.84

0.50–29.76

0.198

2.07

0.20–21.38

0.540

Analysis of 365 Day Follow-Up

The results of univariate and multivariate analyses for the 365 day follow-up period are reported in Tables 7 and 8, with cutoffs for dichotomization chosen from the literature. Over the longer follow-up period, all three electrocardiographic risk variables were strongly associated with cardiovascular death on univariate analysis. SE-MD showed the highest hazard ratio on univariate analysis (HR 7.78, p < 0.001) and was the only variable associated with death on multivariate analysis (HR 4.89, p = 0.003). However, the hazard ratios obtained in this case were lower than the results obtained earlier. The c-statistic (0.78) was also less than the value obtained for the 90 day follow-up (0.85).This suggests that while SE-MD is still a good risk stratifier for cardiovascular death over the year following NSTEACS, it is particularly well suited for the period immediately following the index event.
Table 7

Discriminative ability of SE-MD, HRV and DC measured using the c-statistic to predict cardiovascular death over 365 days following NSTEACS (n = 600)

Parameter

c-Statistic

SE-MD

0.78

HRV

0.66

DC

0.73

Table 8

Univariate and multivariate association between risk variables and cardiovascular death over 365 days following NSTEACS (n = 600)

Parameter

Univariate hazard ratio

95% confidence interval

p value

Multivariate hazard ratio

95% confidence interval

p value

SE-MD

7.78

3.17–19.09

<0.001

4.89

1.70–14.08

0.003

HRV

3.77

1.58–8.99

0.003

1.50

0.55–4.11

0.432

DC

6.13

2.57–14.62

<0.001

2.26

0.85–6.03

0.104

Summary and Discussion

This manuscript investigates use of a new method to risk stratify patients for future cardiovascular death by measuring morphologic differences in ECG affecting the entire heart beat signal. Our method uses dynamic time-warping to compute time-aligned morphology changes between consecutive sinus beats, and then estimates energy at frequencies between 0.30 and 0.55 Hz in the resulting time series of morphologic differences.

We developed the spectral energy of morphologic differences (SE-MD) metric on 764 patients from the TIMI DISPERSE2 study. The SE-MD metric was then evaluated for value in predicting patients at risk of cardiovascular death following NSTEACS in a study using previously unseen data from the TIMI MERLIN study. On a test population of 600 patients from the MERLIN study, SE-MD was strongly associated with the endpoint of cardiovascular death over a 90 day follow-up period. This relationship was consistent for a longer follow-up period of 1 year, and when all risk variables were dichotomized into high risk quartiles or into groups matching the rate of death in the TIMI DISPERSE2 dataset.

For the endpoint of cardiovascular death, SE-MD showed a higher c-statistic and hazard ratio than both HRV and DC for the different experiments considered. We note that these findings hold on a specific endpoint and test population (i.e., cardiovascular death in patients following NSTEACS), and caution against interpreting the results as a general comparison of the different electrocardiographic risk variables. We further observe that a more comprehensive comparison of the methods would explore the risk stratification utility of the different risk variables over periods longer than the 1 year maximum considered by our work and for other endpoints, e.g., MI.

The decision to calculate SE-MD, HRV and DC on 128 Hz holter ECG signals also represents a practical use-case of these methods, i.e., the ability to compute prognostic parameters on data that is typically available for all admitted patients. With higher sampling rates the performance of these metrics could potentially be improved.

A further limitation of this study is the omission of the significant body of work related to the analysis of T-wave morphology (Acar et al. 1999; Zabel et al. 2000). In particular, T-wave alternans (TWA) (Rosenbaum et al. 1994) is commonly used as a measure of repolarization abnormalities that may help identify patients at increased risk. The measurement of TWA typically requires the use of specialized equipment and maneuvers to increase heart rate. We were therefore unable to measure TWA on the holter recordings available in the MERLIN study for comparison with SE-MD. However, we observe that our work differs from TWA both in terms of using information that can be obtained directly from standard holter recordings, and in that SE-MD does not rely on changes in any particular segment of the ECG signal.

Finally, we observe that while we used the endpoint of death due to cardiovascular causes in the reported study, we did not have a more fine-grained description of the cause of death available to us. We believe that death in the period following NSTEACS is likely to be the result of fatal arrhythmias, and SE-MD may help identify patients in a proarrhythmic state. This would be consistent with the results in “Analysis of 365 Day Follow-up” showing SE-MD to be particularly well-suited for risk stratification immediately following the index event. However, we were unable to prove this hypothesis on the current dataset. For future work, we hope to explore the use of SE-MD on a larger cohort, with a more precise categorization of cardiovascular death and a measure of risk variables such as TWA obtained through specialized gold-standard tests.

Notes

Acknowledgments

We would like to thank Gari Clifford for providing tools from Physionet for preprocessing ECG signals and for his technical input on heart rate variability metrics, and Dorothy Curtis for helping with the computational needs for this project. This work was supported, in part, by the Center for Integration of Medicine and Innovative Technology (CIMIT), the Harvard-MIT Division of Health Sciences and Technology (HST), and the Industrial Technology Research Institute (ITRI).

References

  1. Acar B, Yi G, Hnatkova K, Malik M. Spatial, temporal and wavefront direction characteristics of 12-lead T-wave morphology. Med Biol Eng Comput. 1999;37:574–84.PubMedCrossRefGoogle Scholar
  2. Bauer A, Kantelhardt JW, Barthel P, Schneider R, Ing D, Makikallio T, et al. Deceleration capacity of heart rate as a predictor of mortality after myocardial infarction: cohort study. Lancet. 2006;367:1674–81.PubMedCrossRefGoogle Scholar
  3. Ben-Haim SA, Becker B, Edoute Y, Kochanovski M, Azaria O, Kaplinksy E, et al. Beat-to-beat electrocardiographic morphology variation in healed myocardial infarction. Am J Cardiol. 1991;68:725–8.PubMedCrossRefGoogle Scholar
  4. Cannon CP, Husted S, Harrington RA, Scirica BM, Emanuelsson H, Peters G, et al. Safety, tolerability, and initial efficacy of AZD6140, the first reversible oral adenosine diphosphate receptor antagonist, compared with clopidogrel, in patients with non-ST-segment elevation acute coronary syndrome. J Am Coll Cardiol. 2007;50:1844–51.PubMedCrossRefGoogle Scholar
  5. Clifford G, Tarassenko L. Quantifying errors in spectral estimates of HRV due to beat replacement and resampling. IEEE Trans Biomed Eng. 2005;52:630–8.PubMedCrossRefGoogle Scholar
  6. Cox DR, Oakes D. Analysis of survival data. London: Chapman and Hall; 1984.Google Scholar
  7. DeChazal P, O’Dwyer M, Reilly R. Automatic classification of heartbeats using ECG morphology and heartbeat interval features. IEEE Trans Biomed Eng. 2004;51:1196–206.CrossRefGoogle Scholar
  8. Donoho D. Denoising by soft-thresholding. IEEE Trans Inf Theory. 2005;41:613–27.CrossRefGoogle Scholar
  9. El-Sherif N, Hope RR, Scherlag BJ, Lazara R. Reentrant ventricular arrhythmias in the late myocardial infarction period. Circulation. 1977;55:702–19.PubMedGoogle Scholar
  10. Hamilton PS. Open source ECG analysis software documentation. Somerville: EP Limited; 2002.Google Scholar
  11. Hamilton PS, Tompkins WJ. Quantitative investigation of QRS detection rules using the MIT/BIH arrhythmia database. IEEE Trans Biomed Eng. 1986;33:1157–65.PubMedCrossRefGoogle Scholar
  12. Josephson ME, Wit AL. Fractionated electrical activity and continuous electrical activity: fact or artifact? Circulation. 1984;70:529–32.PubMedGoogle Scholar
  13. Li Q, Mark RG, Clifford GD. Robust heart rate estimate fusion using signal quality indices and a Kalman filter. Physiol Meas. 2008;29:15–32.PubMedCrossRefGoogle Scholar
  14. Lomb NR. Least-squares frequency analysis of unequally spaced data. Astrophys Space Sci 1976;39:446–62.CrossRefGoogle Scholar
  15. Malik M. Heart rate variability: standards of measurement, physiological interpretation, and clinical use. Circulation. 1996;93:1043–65.Google Scholar
  16. Morrow DA, Scirica BM, Prokopczuk E, Murphy SA, Budaj A, Varshavsky S, et al. Effects of ranolazine on recurrent cardiovascular events in patients with non-ST-elevation acute coronary syndromes: the MERLIN-TIMI 36 randomized trial. JAMA. 2007;297:1775–83.PubMedCrossRefGoogle Scholar
  17. Myers C, Rabiner L, Rosenberg A. Performance tradeoffs in dynamic time-warping algorithms for isolated word recognition. IEEE Trans Acoust. 1980;6:623–35.CrossRefGoogle Scholar
  18. Ohman E, Granger C, Harrington R, Lee K. Risk stratification and therapeutic decision-making in acute coronary syndromes. JAMA. 2000;284:876–8.PubMedCrossRefGoogle Scholar
  19. Rabiner L. Considerations in dynamic time-warping algorithms for discrete word recognition. IEEE Trans Signal Process. 1978;26:575–82.CrossRefGoogle Scholar
  20. Rosenbaum DS, Jackson LE, Smith JM, Garan H, Ruskin JN, Cohen RJ. Electrical alternans and vulnerability to ventricular arrhythmias. N Engl J Med. 1994;330:235–41.PubMedCrossRefGoogle Scholar
  21. Syed Z, Guttag J, Stultz C. Clustering and symbolic analysis of cardiovascular signals: discovery and visualization of medically relevant patterns in long-term data with limited prior knowledge. EURASIP J Adv Signal Process. 2007;2007:90. doi:10.1155/2007/67938.CrossRefGoogle Scholar
  22. Zabel M, Acar B, Klingenheben T, Franz MR, Hohnloser SH, Malik M. Analysis of 12-lead T-wave morphology for risk stratification after myocardial infarction. Circulation. 2000;102:1252–7.PubMedGoogle Scholar
  23. Zong W, Moody GB, Jiang D. A robust open-source algorithm to detect onset and duration of QRS complexes. Comput Cardiol. 2003;30:737–40.CrossRefGoogle Scholar

Copyright information

© Springer Science+Business Media, LLC 2009

Authors and Affiliations

  • Zeeshan Syed
    • 1
    • 2
  • Phil Sung
    • 1
  • Benjamin M. Scirica
    • 3
  • David A. Morrow
    • 3
  • Collin M. Stultz
    • 1
    • 2
  • John V. Guttag
    • 1
  1. 1.Department of Electrical Engineering and Computer Science, MITCambridgeUSA
  2. 2.Harvard-MIT Division of Health Sciences and TechnologyCambridgeUSA
  3. 3.TIMI Study Group, Cardiovascular DivisionBrigham and Women’s HospitalBostonUSA

Personalised recommendations