# Deep conditional generative model for personalization of 12-lead electrocardiograms and cardiovascular risk prediction

## TL;DR

This paper introduces a deep learning model that generates personalized ECGs and predicts cardiovascular risk by incorporating demographic and anatomical data.

## Contribution

The novel contribution is a conditional VAE framework that synthesizes realistic, subject-specific ECGs and improves cardiovascular risk prediction using latent embeddings.

## Key findings

- The cVAE model generates physiologically consistent ECGs by incorporating demographic and anatomical metadata.
- The revised Cox model achieves a C-index of 0.65 for cardiovascular risk prediction using ECG-derived latent embeddings.
- Generated ECGs show strong consistency with in silico simulations, revealing insights into cardiac anatomy and ECG morphology.

## Abstract

12-lead electrocardiograms (ECGs) are a cornerstone for diagnosing and monitoring cardiovascular diseases (CVDs). They play a key role in detecting abnormalities such as arrhythmias and myocardial infarction, enabling early intervention and risk stratification. However, traditional analysis relies heavily on manual interpretation, which is time-consuming and expertise-dependent. Moreover, existing machine learning models often lack personalization, as they fail to integrate subject-specific anatomical and demographic information. Advances in deep generative models offer an opportunity to overcome these challenges by synthesizing personalized ECGs and extracting clinically relevant features for improved risk assessment.

We propose a conditional Variational Autoencoder (cVAE) framework to generate realistic, subject-specific 12-lead ECGs by incorporating demographic metadata, anatomical heart features, and ECG electrodes’ positions as conditioning factors. This allows for physiologically consistent and personalized ECG synthesis. Furthermore, we introduce a revised Cox proportional-hazards regression model that utilizes the latent embeddings learned by the cVAE to predict future CVD risk. This approach not only enhances the interpretability of ECG-derived risk factors but also demonstrates the potential of deep generative models in personalized cardiac assessment.

Our model is trained and validated on the UK Biobank dataset and in silico simulation data. By incorporating heart position and electrodes’ positions, the generated ECGs demonstrate strong consistency with in silico simulations, providing insights into the relationship between cardiac anatomy and ECG morphology. Furthermore, our CVD risk prediction model achieves a C-index of 0.65, indicating that ECG signals, together with demographic and anatomical information, contain valuable prognostic information for stratifying subjects based on future cardiovascular risk.

This work marks a significant advancement in ECG analysis by providing a conditional VAE framework that not only improves ECG generation but also enriches our understanding of the relationship between ECG patterns and subject-specific information. Importantly, our approach enables clinically significant information to be extracted from 12-lead ECGs, providing valuable insights for predicting future CVD risks.