Dynamic risk assessment of severe ventricular arrhythmia in patients with implantable defibrillators and reduced left ventricular ejection fraction

Background Implantable cardioverter-defibrillator (ICD) shocks have important prognostic implications in patients with polymorphic ventricular tachycardia (PVT) or ventricular fibrillation (VF) events. Purpose This study aimed to develop and validate a predictive model based on dynamic changes of intracardiac electrograms that allows for risk assessment of PVT or VF events in patients with ICDs and heart failure with reduced ejection fraction (HFrEF). Methods The study included a group of pigs with HFrEF (N=8) to characterize the intracardiac and surface electrical changes associated with an increase in cardiac intracavitary pressures during HF decompensation. Sham-operated animals (N=6) were used as comparative controls (Figure 1). In the clinic, the study used clinical and remote monitoring data from two independent cohorts of patients with ICDs. The first cohort (5089 patients, 51 centers and 199,218 tracings) was used for training and internal validation of a convolutional neural network (CNN) model that incorporated electrical parameters associated with HF decompensation and allowed for dynamic risk prediction of a first PVT or VF event. The second cohort included patients implanted with a different ICD vendor (2040 patients, 7 centers and 369,845 tracings) and was used for external validation. Results In pigs, HF decompensation was associated with a prolongation of the QRS complex on the surface ECG and a decrease in R-wave amplitude values of the local bipolar electrograms from the ICD lead. Hemodynamic and electrical alterations of HFrEF decompensation significantly increased the risk of inducible VF episodes compared with sham-operated controls (87.5% vs 0%, respectively; p<0.001). In the clinic, CNN-derived prediction of a first VF or PVT event using the temporal patterns of R-wave amplitude values and QRS complex duration, and additional premature ventricular complex burden, provided receiver operating characteristic areas under the curve (AUROC) of 0.785 (0.757, 0.814) for the training cohort, and 0.725 (0.709, 0.739) for the external validation cohort (Figure 2). Dynamic risk prediction of the CNN model outperformed the predictive metrics of static risk assessment using clinical variables (AUROC 0.616 [0.601, 0.632] for external validation) (Figure 2). Conclusions A CNN-derived model, grounded in the pathophysiology of HFrEF, provides dynamic risk assessment of VF or PVT events in patients ICD devices.