Abnormal repolarization in the acute myocardial infarction patients: a frequency-based characterization.

Despite ST elevation having poor sensitivity for acute myocardial infarction (AMI), it remains the main electrocardiographic (ECG) repolarization index for AMI diagnosis. Aim of the present study was to propose a new f99 index, defined as the frequency at which the repolarization normalized cumulative energy reaches 99%, for ECG AMI discrimination from health with good sensitivity and good specificity. Evaluation of such f99 index was performed on 12-standard-lead (I, II, III, aV1, aVr, aVf, V1 to V6) ECG recordings of 47 healthy controls and 108 acute myocardial infarction (AMI) patients. Repolarization dispersion caused f99 distributions to be significantly lead dependent. In most leads (leads I, II, aVl, aVr, V2-V6), f99 median value was lower in the healthy controls (10-17 Hz) than in the AMI patients (12-38 Hz) indicating higher frequency components (i.e. a more fragmented repolarization) in the latter population. AMI patients from healthy controls discrimination by f99, evaluated in terms of sensitivity (Se) and specificity (Sp), was also lead dependent. Single-lead analysis indicated leads I (Se=80%, Sp=77%) and aVl (Se=84%, Sp=74%) as optimal. Instead, lead-system analysis, performed to overcome dispersion issues, provided the best results when averaging over the 6 precordial leads (Se= 81% and Sp=74%). In conclusion, our new f99 index appears as a promising tool for non-invasively and reliably discriminate AMI patients from healthy subjects.

INTRODUCTION

Defects in the cardiac repolarization are known to be associated to several life-threatening diseases [1-4]. In the electrocardiogram (ECG) such defects appear as abnormalities of the ST segment and T-wave, which can be non-invasively characterized by means of indexes. The most popular ECG repolarization indexes are the QT interval [2,3], the ST elevation [5,6] and the T-wave alternans [4,7]. Additional morphological indexes present in the literature include the T-wave duration parameters [8,9], the T-wave amplitude parameters [10,11] and others [12-16]. Abnormalities in the ECG repolarization morphology are reflected, in the frequency domain, in a variation of the T-wave frequency content. Though, repolarization indexes based on this feature have only occasionally been proposed [17-19].

The acute myocardial infarction (AMI: a disease in which the blood stops flowing properly to part of the heart and the heart is injured because not receiving enough oxygen) is one of the leading causes of death and disability in the world [20]. Even though in most cases ST elevation is a result of non-AMI causes [5] and several other ECG repolarization abnormalities have been observed in the AMI patients [2,18, 19,21,22], the ST elevation remains the ECG repolarization index mainly used to diagnose the presence of an AMI [23,24] since initial ST elevation is part of the classic evolutionary pattern of the AMI [25]. Still, ST elevation has poor sensitivity for AMI given that up to 50% of patients exhibit atypical repolarization changes which may include isolated ST depression, T inversion or even normal ECG [24]. In order to significantly improve the rate of correct AMI diagnosis from ECG tracings, new repolarization parameters are needed. Thus, aim of the present study was to propose a new repolarization index based on the T-wave frequency content, termed f99, able to discriminate AMI patients from healthy subjects with both a good sensitivity and a good specificity. Being defined in the frequency domain, f99 computation requires automatic analysis of the ECG signals. Often, many of the above-mentioned time-domain repolarization parameters are still manually evaluated. Compared to human evaluation, automatic ECG analysis allows a more objective characterization of the ECG features by eliminating the subject-related variations of the measures, even though some inter-method variations may persist [26,27].

METHODS

Clinical Data

Our clinical data consisted of short (30 s to 2 min) 12-standard-lead (I, II, III, aV1, aVr, aVf, V1 to V6) digital ECG recordings (1 KHz sampling frequency) from 47 healthy controls and 108 myocardial infarction patients (AMI), all belonging to the PTB Diagnostic Physionet database (www.physionet.org). Among the AMI patients, 49 were affected by anterior acute myocardial infarction (ANAMI) and 59 by inferior acute myocardial infarction (INAMI). Additional clinical information relative to the two populations is reported in Table .

Repolarization Characterization

Preprocessing. Before T-wave frequency-content evaluation, ECG tracings underwent a preprocessing stage consisting of F=200 Hz resampling, low-frequency (≤0.5 Hz) noise removal and 50 Hz line noise removal. Eventually, after R-peak detection, a 20-beat 12-lead ECG window characterized by stable HR (RR-interval standard deviation<10% of mean RR) and no ectopic beats and artifacts was randomly extracted from each ECG tracing.

Repolarization frequency-content characterization. Repolarization frequency-content evaluation was performed in each ECG lead independently. From the median beat, computed using the 20 available beats, the repolarization onset (RepOn) and offset (RepOff) were identified as:

Eq. (2) is an adjustment of an empirical formula [28] finalized to avoid cases of P-wave inclusion in the T-wave window, and median RR (s) is the median RR interval. The median repolarization waveform was then forced to be 260 ms long by opportune resampling. Eventually, the repolarization signal (RPS) was constructed by zero padding everything outside the resampled median repolarization waveform till 1 second.

RPS frequency-content evaluation was performed by computing the Fourier power spectrum (PS; Eq. 3) and energy signal (E; Eq. 4):

where N is the number of samples (N=200), and n and k are adimensional indexes to get time and frequency as t=n·(1/F)=n·0.005 s, with n=1, 2, …N, and f=k Hz, with k=1, 2, …N/2, respectively. After having computed the total energy (E; Eq. 5), the PS and the E were normalized and expressed as percentages (PS and E, respectively; Eq. 6 and Eq. 7):

By definition, E is a monotonically increasing function which saturates at 100%. The frequency at which E first reaches or overcomes 99%, called f99, represents an index to characterize repolarization.

Abnormal repolarization identification. Abnormal repolarization was identified when f99 exceeded a threshold defined as the 75th percentile of the f99 distribution over the healthy controls.

Repolarization Analysis Types

Two types of repolarization analysis were performed:

Single-lead analysis. The f99 values were computed in each ECG lead, and the abnormal repolarization identification criterion was then applied to their single-lead distributions.

Lead-system analysis. The single-lead f99 values were averaged over the 6 precordial and 12-standard-leads, and the abnormal repolarization identification criterion was then applied to the distributions of the averaged f99 values.

Algorithm Implementation and Simulation Test

The f99 algorithm was implemented in MATLAB and validated using simulated data with known input data. Specifically, a signal y(t), constituted by the summation of two sinusoids y1(t) and y2(t), was considered:

Validation was performed by keeping constant the values of f, f, and A(f=1 Hz; f=15 Hz; A=1 a.u.), while the A was varying from 0 to 1 a.u. (0.01 steps). Being f99 the frequency at which the signal total energy (E=E+E=A2+ A2) reaches or overcomes 99%, f99 is correctly estimated if it equals 1 Hz till A =A* and then becomes equal to 15 Hz. A* is obtained by putting E=0.99·(E+E), which implies 0.99·(A*)2 = (1-0.99) A2 , that is:

Robustness Tests

To evaluate f99 robustness, three kinds of tests were considered:

Robustness to dispersion. Single-lead f99 distributions were compared to evaluate the presence of f99 inter-lead dispersion.

Robustness to repolarization window offset. The f99 values were computed using the nominal RepOn and RepOff points (Eq. 1 and Eq. 2) and after a random ±20 ms shift of RepOff from its nominal location. Results obtained in the two experiments were compared to investigate f99 dependency on RepOff identification.

Robustness to heart rate. The association between f99 and corresponding HR was evaluated to investigate the dependency of the former parameter on the latter.

Statistics

Normality of parameters distributions was tested using the Lilliefors test. Parameters characterized by normal and not-normal distributions were compared using the T-test and the Wilcoxon rank-sum test, respectively. Binary parameters distributions were compared using the chi-square test or, when not possible (expected cell frequency <5), the Fisher exact one-tailed probability test. The Kruskal-Wallis test was used to perform the one-way ANOVA test to compare distributions over the leads. Lead-pairs comparison was performed using the multiple comparison procedure. Associations between two parameters distributions were evaluated using the correlation coefficient (ρ). Eventually, f99 ability to discriminate abnormal repolarization was evaluated using sensitivity (Se), and specificity (Sp, fixed at 75% by the abnormal repolarization criterion defined above). Statistical significance level was 0.05.

RESULTS

Simulation study. In our simulation study, f99 resulted equal to 1 Hz for A ranging from 0 to A*=0.1, and to 15 Hz for higher values of A (Fig. ), confirming the goodness of our f99-algorithm MATLAB implementation.

Single-lead analysis. In most leads, the AMI patients’ spectra were characterized by larger high-frequency components than the healthy controls’ spectra (Fig. ). Consequently, compared to the AMI patients, the healthy controls were usually characterized by E curves saturating at lower frequencies (Fig. ) and by significantly lower f99 values (Table ). Lead III was the only one showing an opposite trend, whereas aVf and V1 f99 distributions did not differ significantly between the two populations (Table ). Analogous findings were obtained when comparing ANAMI and INAMI subgroups against the healthy controls (Table ). Ability of f99 in discriminating pathological from healthy conditions, evaluated in terms of Se and Sp, was lead dependent (Table ). Optimal leads could also vary with AMI subgroup. Leads I and aVl discriminated both the ANAMI and INAMI patients from the healthy controls. Instead, lead V2 discriminated the ANAMI group better than the INAMI group, while lead V5 vice versa (Table ).

Lead-system analysis. Mean f99 values, averaged over the 6 precordial and 12-standard-leads, were significantly lower in the healthy controls than the AMI patients (Table ). However, discrimination between the two populations was superior when averaging over the 6 precordial leads than over the 12-standard-leads (Table ). Analogous results were obtained when analyzing the two AMI subgroups independently (Tables and ).

Robustness to dispersion. No specific lead showed significantly higher or lower f99 values compared to all other leads (Table ). However, statistically significant differences were identified among couples of leads, so that final values of Se and Sp were indeed significantly lead dependent (Table ).

Robustness to repolarization window offset. Distributions of f99 parameter were very robust to changes in the RepOff identification in all populations (Table ). Consequently, values of Se and Sp were very close to those obtained using the RepOff nominal value (Table ).

Robustness to heart rate. In all populations, f99 values were substantially independent from HR (Table ).

DISCUSSION

This study proposes f99 as an innovative frequency-based index for repolarization characterization and AMI patients discrimination from healthy subjects. The algorithm for f99 computation was tested on simulated data to confirm its proper functioning. Instead, f99 clinical usefulness was evaluated on ECG tracings from 47 healthy controls and 108 AMI patients (49 ANAMI e 59 INAMI), who are known to show abnormal repolarization [17, 21, 22].

The f99 index, defined as the frequency at which the repolarization normalized cumulative energy reaches 99%, was computed in each one of the 12-standard-leads independently, and averaged over the 6 precordial and 12-standard-leads. In most leads (9 out of 12, Table ), f99 median values was significantly lower for the healthy controls than for the AMI patients. This finding implies that AMI T-waves have a normalized power spectrum characterized by higher frequency components which indicate a more fragmented repolarization. On this basis, f99 was used to discriminate abnormal from normal repolarization. More specifically, abnormal repolarization was identified when f99 exceeded a threshold defined as the 75th percentile of the f99 distribution over the healthy controls. Such threshold definition implies a nominal Sp of 75%. However, actual values of Sp were either slightly lower (74%, Table ) or higher (77%-87%, Table ). Lower values of Sp are due to the fact that the healthy-control population counts less than 100 subjects. Instead, higher values of Sp are due to the fact that there can be more than one healthy control characterized by the same f99 value immediately under threshold. Se values were strongly lead dependent (Se=6%-85%, Table ), indicating a significant repolarization dispersion identified by f99. Leads I and aVl were able to significantly discriminate abnormal from normal repolarization, independently of the infarct location (lead I: Se=78%-82%, Sp=77%; lead aVl: Se=84%-85%, Sp=74%, Table ). Instead, lead V2 allowed identification of abnormal repolarization only in the ANAMI patients (Se=82%, Sp=79%, Table ) while lead V5 only in the INAMI patients (Se=78%, Sp=79%, Table ). These findings indicate that f99 is characterized by a certain amount of dispersion, similarly to what happens for other repolarization indexes [29,30]. Since the leads provide spatial information about the heart's electrical activity in different directions, the optimal lead for discriminating abnormal repolarization by f99 may depend on the infarct location. More specifically, leads closer to the AMI location will identify the disease better than the others.

Since it is not possible to know a priori which will be the optimum lead to identify an AMI, to overcome issues related to repolarization dispersion, values of f99 averaged over the 6 precordial and 12-standard-leads were provided. Between these two averaged indexes, the former discriminated abnormal repolarization (Se= 80%-81%, Sp=74%, Table ) better than the latter (Se=69%, Sp=74%, Table ) since the f99 distributions were showing a more homogeneous trend among the precordial leads than the limb and augmented-limb leads (Table ). Thus, lead-system analysis showed that the 6 precordial leads were superior to the 12-standard-leads in discriminating abnormal repolarization, even though the single-lead analysis showed that the optimal lead could not be precordial.

The f99 index, either measured on a single lead or averaged over a lead system, is statistically independent from the exact determination of the repolarization window (Tables and ), and does not require the exact identification of the T-wave offset, usually very difficult in clinical cases [9]. T-wave offset estimation can indeed occur with a few tens of ms of tolerance (Table ). In addition, f99 was independent from HR (Table ). In general, repolarization significantly depends on HR [10] so that some indexes, like the popular QT, may require a correction to adjust to it. The f99 independency of HR is due to the fact that such adjustment is performed during the construction of the RPS (see Methods), which requires a stretch or a compression of the repolarization segment to match a length of 260 ms (assumed as standard), and a subsequent zero padding to reach a 1 s length for each heart cycles. Thus, the normalized power spectrum does not provide the frequency content of the real T-wave, but rather the amplitude of the harmonics after forcing the fundamental frequency to be at 1 Hz. In other words, f99 although expressed in Hz for simplicity and clarity, represents the number of harmonics the normalized cumulative energy needs to reach 99%.

CONCLUSION

The present study proposed a new f99 index, defined as the frequency at which the repolarization normalized cumulative energy reaches 99%, for ECG AMI discrimination from health. Evaluation of such f99 index was performed on 12-standard-lead ECG recordings of 47 healthy controls and 108 AMI patients. Repolarization dispersion caused f99 distributions to be significantly lead dependent indicating that optimal lead for discriminating abnormal repolarization may depend on the infarct location. To overcome dispersion issues, lead-system analysis was performed by averaging f99 over the 6 precordial leads and proved to be able to identify the presence of AMI with good sensitivity and specificity. Thus, our new f99 index appears as a promising tool for non-invasively and reliably discriminate AMI patients from healthy subjects.

Table 1.

Clinical parameters. Mean±standard deviation values or number of occurrences are reported.

	Healthy Controls (47)	ANAMI (49)	INAMI (59)	AMI (108)
Age (year)	45±15	60±11°	58±11°	59±11°
Gender (male)	37 (82%)	39 (80%)	49 (83%)	88 (82%)
Hypertension	0 (0%)	12 (25%)§	20 (34%)°	32 (30%)°
Obesity	0 (0%)	4 (8%)	5 (8%)*	9 (8%)
Diabetes	0 (0%)	4 (8%)	12 (20%)§	16 (15%)§
Other pathologies	0 (0%)	28 (57%)°	29 (49%)°	57 (53%)°
Heart rate (bpm)	69±12	82±15°	81±17§	81±16°

P<0.05 when comparing against the healthy controls

P<10-3 when comparing against the healthy controls

P<10-5 when comparing against the healthy controls

Table 2.

Values of  50th [25th, 75th] percentiles of the f99 distributions.

	Healthy Controls (47)	ANAMI (49)		INAMI (59)		AMI (108)
	f99 (Hz)	f99 (Hz)	P	f99 (Hz)	P	f99 (Hz)	P
Single-lead analysis
I	13 [12,15]	40 [23,54]	<10-7	34 [18,46]	<10-7	37 [20,52]	<10-9
II	17 [13,24]	34 [15,53]	<10-3	38 [17,53]	<10-3	37 [15,53]	<10-4
III	50 [22,60]	23 [11,44]	<10-4	25 [14,43]	<10-3	24 [13,43]	<10-4
aVl	13 [12,18]	42 [22,54]	<10-9	35 [21,53]	<10-8	38 [22,53]	<10-11
aVr	15 [11,46]	26 [12,50]	NS	27 [15,46]	<0.05	27 [15,50]	<0.05
aVf	24 [14,50]	24 [14,39]	NS	22 [12,40]	NS	22 [13,40]	NS
V1	18 [13,25]	23 [15,31]	NS	12 [10,41]	NS	19 [11,35]	NS
V2	10 [9,10]	15 [12,22]	<10-7	10 [9,12]	NS	12 [10,20]	<10-4
V3	10 [9,11]	14 [11,20]	<10-8	11 [10,21]	<10-3	13 [10,21]	<10-6
V4	11 [10,12]	14 [11,22]	<0.01	19 [11,38]	<10-4	15 [11,29]	<10-4
V5	13 [11,15]	25 [12,43]	<0.01	40 [18,53]	<10-7	31 [14,47]	<10-6
V6	14 [13,22]	37 [18,53]	<10-5	36 [17,53]	<10-4	37 [18,53]	<10-6
Lead-system analysis
Mean V1-V6	14 [13,16]	24 [16,31]	<10-7	25 [17,31]	<10-6	24 [17,31]	<10-8
Mean 12 std	19 [16,23]	28 [22,34]	<10-5	30 [21,35]	<10-5	29 [21,35]	<10-6

P-value when comparing f99 values against the healthy controls

not statistically significant (P≥0.05)

Table 3.

Values of f99 threshold over which abnormal repolarization is recognized together with corresponding values of sensitivity (Se) and specificity (Sp).

		Healthy Controls	ANAMI	INAMI	AMI
	f99 threshold (Hz)	Sp	Se	Se	Se
Single-lead analysis
I	15.0	77%	82%	78%	80%
II	24.3	74%	61%	61%	61%
III	60.0	77%	6%	7%	6%
aVl	17.8	74%	84%	85%	84%
aVr	45.5	74%	33%	25%	29%
aVf	50.3	74%	16%	17%	17%
V1	25.0	77%	43%	32%	37%
V2	10.0	79%	82%	49%	64%
V3	11.0	87%	71%	44%	56%
V4	12.0	83%	57%	63%	60%
V5	15.0	79%	57%	78%	69%
V6	22.0	79%	65%	69%	68%
Lead-system analysis
Mean V1-V6	15.9	74%	80%	81%	81%
Mean 12 Std	23.1	74%	69%	69%	69%

Table 4.

Values of the correlation coefficient (ρ) between f99 distributions obtained using nominal and randomly shifted (within ±20 ms) RepOff points.

	Healthy Controls	ANAMI	INAMI	AMI
Single-lead analysis
I	0.98	0.97	0.99	0.98
II	0.99	0.98	0.99	0.99
III	0.98	0.99	0.99	0.99
aVl	0.98	0.98	0.99	0.99
aVr	1.00	0.99	0.99	0.99
aVf	0.99	0.99	0.99	0.99
V1	0.99	0.99	1.00	1.00
V2	0.99	0.95	0.98	0.97
V3	0.96	0.94	0.100	0.98
V4	0.99	0.94	0.99	0.98
V5	0.99	0.97	0.99	0.98
V6	0.99	0.97	0.99	0.98
Lead-system analysis
Mean V1-V6	0.98	0.97	0.99	0.98
Mean 12 Std	0.99	0.98	0.99	0.99

Table 5.

Values of f99 threshold over which abnormal repolarization is recognized together with corresponding values of sensitivity (Se) and specificity (Sp) when RepOff was randomly shifted within  ±20 ms from its nominal value.

		Healthy Controls	ANAMI	INAMI	AMI
	f99 threshold (Hz)	Sp	Se	Se	Se
Single-lead analysis
I	16.0	77%	84%	80%	81%
II	23.5	74%	63%	59%	61%
III	59.0	77%	6%	10%	8%
aVl	18.8	74%	84%	83%	83%
aVr	42.8	74%	33%	27%	30%
aVf	51.3	74%	16%	17%	17%
V1	25.5	74%	45%	34%	39%
V2	10.0	77%	82%	49%	64%
V3	11.0	81%	65%	47%	56%
V4	12.0	79%	61%	61%	61%
V5	15.8	74%	59%	80%	70%
V6	21.0	77%	67%	71%	69%
Lead-system analysis
Mean V1-V6	16.0	77%	78%%	81%	80%
Mean 12 Std	23.4	74%	71%	73%	72%

Table 6.

Values of the correlation coefficient (ρ) between f99 distributions and heart rate.

	Healthy Controls	ANAMI	INAMI	AMI
Single-lead analysis
I	0.28	0.44*	0.07	0.23*
II	0.42*	0.41*	0.11	0.22*
III	0.08	0.41*	0.12	0.24*
aVl	0.46*	0.37*	0.16	0.24*
aVr	0.14	0.36*	0.04	0.18
aVf	0.31*	0.52*	0.09	0.26*
V1	0.03	0.14	-0.05	0.01
V2	-0.02	0.20	-0.12	-0.02
V3	0.18	0.20	0.10	0.13
V4	0.11	0.19	0.25	0.21*
V5	0.20	0.35*	0.07	0.16
V6	0.34*	0.18	-0.06	0.04
Lead-system analysis
Mean V1-V6	0.20	0.33*	0.05	0.14
Mean 12 Std	0.36*	0.56*	0.12	0.29*