Long-term prediction of heart failure using an ECG-based fine-tuned self-supervised neural network model identifies 23 genetic loci implicating electrophysiological mechanisms

Background Risk stratification of heart failure (HF) in individuals without prevalent heart disease is crucial for early intervention and prevention. Electrocardiogram (ECG)-based artificial intelligence models have shown promise when predicting HF events at long follow-up periods. However, they are still limited by imbalanced case-control data and small labelled datasets. Moreover, the electrophysiological mechanisms underlying HF remain understudied. Purpose We compared the performance of a baseline convolutional neural network (CNN) model designed to predict HF events from the ECG at 8 different timepoints (1, 2, 3, 4, 6, 8, 10 & 12 years after the ECG acquisition), with a fine-tuned self-supervised model, pre-trained using a large amount of unlabelled ECGs. Subsequently, we investigated the genetic architecture of the prediction from the strongest model. Methods The baseline model consisted of 6 convolutional layers, an attentive pooling layer and 3 fully connected layers. The fine-tuned self-supervised model was composed of a frozen self-supervised feature extraction block, and a trainable classification block. The self-supervised block, composed of 6 convolutional layers, aimed to extract relevant features from the ECG and was trained with 637,438 publicly available ECGs. The classification block, composed of 1 convolutional layer, an attentive pooling layer and 3 fully connected layers, used a loss function designed to optimize the prediction of HF at the 8 timepoints. The models were trained and tested using 86,567 ECGs from UK Biobank (69,257 for training and 17,310 for test). Then, we performed a genome-wide association study (GWAS) on the standardized output at 8 years of the model with the highest area under the curve (AUC), with sex, age, body mass index, genetic array and the RR interval. Results The fine-tuned self-supervised model showed a better performance across all follow-up times when compared to the baseline model, increasing the AUC 7% on average (Figure 1). The GWAS revealed 27 candidate genes at 23 genome-wide significant loci (P < 5 x 10-8), 19 previously proposed for ECG traits, 9 of these for cardiac structural or functional traits (6 for HF), and 4 new (PATJ, BOLL, NATA/PTGIS and CCDC63). We found these genes were significantly enriched in cardiac conduction and regulation of cardiac contraction processes (Figure 2). Conclusions The use of features derived from a pretrained self-supervised model improved the performance of the CNN model, allowing a better prediction of HF at long follow-up periods. Also, the genetics underlying the ECG-based prediction of HF indicated candidate genes previously reported for ECG and arrhythmia phenotypes, highlighting electrophysiological mechanisms implicated in HF. Our findings support the integration of deep learning with genetic analyses to develop more accurate and clinically valuable risk prediction tools for early intervention and prevention of HF.