The arrhythmic substrate of hypertrophic cardiomyopathy using ECG imaging

Introduction: Patients with hypertrophic cardiomyopathy (HCM) are at risk for lethal ventricular arrhythmia, but the electrophysiological substrate behind this is not well-understood. We used non-invasive electrocardiographic imaging to characterize patients with HCM, including cardiac arrest survivors. Methods: HCM patients surviving ventricular fibrillation or hemodynamically unstable ventricular tachycardia (n = 17) were compared to HCM patients without a personal history of potentially lethal arrhythmia (n = 20) and a pooled control group with structurally normal hearts. Subjects underwent exercise testing by non-invasive electrocardiographic imaging to estimate epicardial electrophysiology. Results: Visual inspection of reconstructed epicardial HCM maps revealed isolated patches of late activation time (AT), prolonged activation-recovery intervals (ARIs), as well as reversal of apico-basal trends in T-wave inversion and ARI compared to controls (p < 0.005 for all). AT and ARI were compared between groups. The pooled HCM group had longer mean AT (60.1 ms vs. 52.2 ms, p < 0.001), activation dispersion (55.2 ms vs. 48.6 ms, p = 0.026), and mean ARI (227 ms vs. 217 ms, p = 0.016) than structurally normal heart controls. HCM ventricular arrhythmia survivors could be differentiated from HCM patients without a personal history of life-threatening arrhythmia by longer mean AT (63.2 ms vs. 57.4 ms, p = 0.007), steeper activation gradients (0.45 ms/mm vs. 0.36 ms/mm, p = 0.011), and longer mean ARI (234.0 ms vs. 221.4 ms, p = 0.026). A logistic regression model including whole heart mean activation time and activation recovery interval could identify ventricular arrhythmia survivors from the HCM cohort, producing a C statistic of 0.76 (95% confidence interval 0.72–0.81), with an optimal sensitivity of 78.6% and a specificity of 79.8%. Discussion: The HCM epicardial electrotype is characterized by delayed, dispersed conduction and prolonged, dispersed activation-recovery intervals. Combination of electrophysiologic measures with logistic regression can improve differentiation over single variables. Future studies could test such models prospectively for risk stratification of sudden death due to HCM.


Analysis of surface ECG markers
To examine for body surface recording signs of conduction pathology, QRS durations were measured for the peak exercise and recovery datasets.The vest output rather than conventional 12-lead was used to avoid timing issues with the exercise machine output.The positional equivalent of 12-lead ECG V2 was used (electrodes 71-76 on the CardioINSIGHT™ vest), with the first beat from the sample as the representative measurement.

Measures of epicardial electrophysiology
For a given epicardial electrogram, local activation time (AT) was defined as the period from QRS start to steepest negative point of the QRS complex, and local activation-recovery intervals (ARI) as local activation time to steepest positive point of the T wave (Wyatt method 4 ).The Wyatt method is favored by ECGi mapping papers to date [5][6][7][8] .
To search for steep electrical gradients, each electrogram location on the epicardial shell was linked to neighboring locations within a 5mm Euclidean search distance.For each node-neighbor pair, the difference in AT or ARI was divided by the distance between the locations, giving a gradient in milliseconds/millimeter.For each node on the epicardial surface, the mean gradient within a 5mm radius was calculated, and these values were averaged across the epicardial shell to give a wholeheart estimation of steep electrical gradients.
To fully understand the electrophysiology of the three groups, three domains were defined for analysis: 1.The mean of activation or ARI was used to describe overall conduction or repolarization delay.2. The central 95% range of times was used to describe dispersion.3. The mean gradient of activation and ARI times in space was used to detect the presence of steep gradients.

Logistic regression for the description of the arrhythmogenic substrate in HCM
To understand the contribution of different parameters from our panel to the arrhythmogenic substrate in HCM, we built multiple variable logistic models from significant variables.To qualify for inclusion, a measure would have to significantly differentiate HCM VF and HCM volunteers (p<0.05).
Qualifying measures were scaled to the mean and variance of the whole dataset.To improve the ability of the model to predict on unseen data, collinearity was reduced by rejecting one measure of any pair with a Pearson correlation of >0.8 9 .A multiple logistic model was then fitted using Newton's method.Backward stepwise selection was used to reject variables with p>0.15 10,11 .Odds ratios were calculated by exponent of the model coefficients.The predicted probability of an observation falling into the HCM VF group was compared for the true HCM VF group, and the HCM group without previous arrhythmia.

Ability of a multiple logistic model to predict in unseen data
To determine the ability of these logistic models to predict whether a patient was in the HCM or HCM VF group in a wider population, k-folds validation was performed.K-folds validation is used in small datasets because it tests on the entire population (), thereby avoiding the potentially large effect of single outliers in small validation sets 12,13 .Briefly, a subset of patients is reserved for testing (size   , a 'fold'), and the remaining patient data is used to train a logistic regression model.The accuracy of this model is then assessed on the reserved testing group.This is repeated by reserving a new testing group and training another model on the remaining data.Once all  folds are tested, the accuracy results are aggregated to estimate sensitivity and specificity.To ascertain the incremental value of ECG imaging over the surface ECG in identifying VF survivors from the HCM cohort, the logistic regression analysis was repeated using only QRS duration and QTc in both exercise and recovery.The both the balanced accuracy and area under the receiver operating characteristic curve of the 2-variable ECGi model was superior to the surface measures only model.

Surface data model regression results
No The software used for analysis was written by two of the study authors (J.C. and M. S-S.).To determine whether an early stage user of the software could reproduce the measurements, an interoperator reproducibility study was performed.R. G. (see acknowledgements), a first-year cardiology registrar (5 years post qualification as a doctor) was recruited as the trainee.J. C. delivered approximately 2 hours of in-person teaching and the trainee was allowed to experiment making measurements through the software unsupervised.
An early issue raised by the trainee was failure of the neural network to segment the T-wave correctly.It was found that the trainee had set the QRST window of interest to include some of the preceding T wave, leading to the neural network highlighting the area from the end of the last QRST to the end of the target QRST.Care to avoid the previous QRST complex allowed normal function of the software.This provided a useful cue to the trainee for when a QRST selection deviated from the ideal.
Following this learning point, the trainee (Operator 1) performed measurements on 10 HCM patients, blinded to the scores of the present study's author (Operator 2).Spearman's R was >0.9 for both AT and ARI measurements, suggesting that even for a new user of the software, reproducible measurements are possible.
Supplementary figure S1: Comparison of whole heart activation and repolarization metrics immediately after peak exercise and in end recovery between hypertrophic cardiomyopathy (HCM) and a selection of structurally normal heart control groups: (I) fully recovered and revascularized ischaemic VF survivors, IHD VF; (II) patients with benign but symptomatic idiopathic ventricular ectopy, VE; (III) the unaffected relatives of patients with Brugada syndrome, BrS relative.Local activation time (LAT) was defined as the onset of the first epicardial QRS complex to the steepest negative slope of the electrogram-QRS complex.Local repolarization time (LRT) was defined as the onset of the first epicardial QRS complex to the steepest positive slope of the electrogram-T wave.Activation recovery interval (ARI) is the difference between LAT and LRT.Mean time is the average of all LAT/LRT/ARI across the heart.Dispersion is the central 95% range of LAT/LRT/ARI across the heart.Gradient is the whole-heart mean rate of range in LAT/ARI over a 5mm search distance around each epicardial location.of LAT/LRT/ARI across the heart.Gradient is the whole-heart mean rate of range in LAT/ARI over a 5mm search distance around each epicardial location.

Table S1 :
Correlation matrix to detect intervariable dependence.High Pearson correlation between two variables suggests a 1:1 relationship and predisposes models to collinearity.In our study we chose to eliminate one of any pair of variables more with a Pearson correlation >0.8 (high interdependence).In this case, mean activation time in exercise was eliminated (high correlation with mean activation time in recovery).Activation recovery interval, ARI.