Non-ischemic cardiomyopathies represent almost 40% of hospitalized patients with heart failure. These pathologies exhibit high variability in clinical presentation and progression due to their multi-factorial nature. Characterizing a patient’s health state requires gathering information from different sources (MRI images, patient record, ECG, etc.) and aggregating them to obtain a simpler representation. I am working on different models within the family of Multimodal Variational Autoencoders (VAEs) which offer a probabilistic framework for learning joint latent representations from heterogeneous data. In this presentation, I will discuss two architectures for multimodal VAEs: alignment-based and fusion-based approaches. While the alignment model learns a latent space for each modality, the fusion model aggregates distributions from each modality into a unified representation. However, the choice of aggregation method raises theoretical challenges, one of them being computing the Kullback-Leibler divergence between latent distributions.