Abstract 12434: ECG-Only Deep Learning Algorithm Predicts Risk of Arrhythmia in Phospholamban Cardiomyopathy

Introduction Phospholamban (PLN) p.Arg14del mutation carriers are at risk of developing arrhythmogenic cardiomyopathy, but phenotypic penetrance is incomplete and much interpatient variability exists in the risk of malignant ventricular arrhythmias (MVA). Accurate risk stratification allows for timely device implantation which may prevent sudden cardiac death (SCD). The current prediction model is multimodal and uses data from ECG, echocardiography, MRI and 24-hour Holter monitoring. Hypothesis We hypothesized that a deep learning-based approach may allow for accurate risk prediction using only 12-lead ECG data. Methods A total of 679 PLN p.Arg14del mutation carriers without MVA at baseline were identified. MVA was defined as sustained VA, appropriate ICD intervention or (aborted) SCD. Prediction models were developed using Cox regression and validated using a bootstrap-based optimism corrected area under the receiver operating curve (AUC). A deep learning-based variational auto-encoder, trained on 1.1 million ECGs, was used to convert the median beat baseline ECG into its FactorECG. The FactorECG is a compressed version of the ECG which summarizes it into 21 explainable factors. Seven ECG factors that were previously shown to be associated with reduced ejection fraction were included in the model. A comparison to conventional ECG parameters and the current prediction model was performed. Results The deep learning-based ECG-only approach was able to predict MVA in PLN mutation carriers with an AUC of 0.82 [95% CI 0.78 - 0.86], comparable to the current multimodal prediction model (AUC 0.84 [95% CI 0.80 - 0.88]) and outperforming a model based on conventional ECG parameters (low voltage ECG and negative T waves; AUC 0.66 [95% CI 0.59 - 0.73]). Conclusions In conclusion, our deep learning-based algorithm uses only ECG data to accurately predict the occurrence of MVA in PLN mutation carriers, which could allow for easier stratification of patients at risk.