| Summary: | <p>This thesis presents a novel deep learning framework based on Variational Autoencoders (VAEs) for the generation and analysis of 12-lead electrocardiograms (ECGs), and its use for cardiovascular disease (CVD) risk prediction. It focuses on three major aspects:</p>
<p>First, we propose a model for 12-lead ECG generation that is capable of generating realistic 12-lead ECG data. Comparison with existing methods shows that ours produces lower Maximum Mean Discrepancy and more lead information, while maintaining generation diversity. The model provides an interpretable latent space to explain the generation process and allows for a deeper understanding of the relationships between ECG features.</p>
<p>Second, we introduce patient-specific characteristics by extending the VAE to a conditional VAE, incorporating essential subject-specific information such as age, sex, Body Mass Index (BMI), and anatomical properties of heart and torso into the ECG generation process. This integration enables our model to generate ECGs that are representative of subjects with diverse profiles. The generated ECGs demonstrate strong correspondence with established models and in-silico experiments. To the best of our knowledge, our model is the first deep learning method to integrate heart information into ECG generation. This brings the potential to improve personalized medicine and contribute to a better understanding of the relationship between heart anatomy and ECG morphology.</p>
<p>Lastly, we focus on future cardiovascular disease risk prediction by leveraging the latent space derived from the VAE. We develop an additional classifier that effectively stratifies patients into low and high-risk groups for CVD. This finding shows the presence of informative ECG features for CVD prediction and demonstrates our model’s ability to extract these features, offering new insights for early CVD risk detection and prevention.</p>
<p>In summary, this work contributes to the growing body of research exploring the potential of deep learning techniques in electrocardiography, and their implications for personalized medicine, improved diagnostics, and early CVD risk detection. The proposed VAE-based framework serves as a robust and versatile tool for advancing the field of ECG analysis and risk prediction.</p>
|
|---|