<p>Recently, causal analysis of clinical data has become essential for understanding the mechanisms underlying complex diseases. Traditional causal discovery methods, such as LiNGAM, provide interpretable causal graphs but often fail to generate informative latent embeddings for predictive tasks. Meanwhile, deep graph learning techniques, particularly variational graph autoencoders (VGAEs), can learn powerful latent embeddings through encoder–decoder architectures, yet they frequently overlook causal relationships, limiting interpretability and clinical applicability. To address these limitations, we propose a hybrid framework, named GraCa, which integrates causal inference via LiNGAM with VGAE-based deep graph learning. In this approach, the causal graph estimated by LiNGAM serves as a structural prior to guide both the encoder and decoder, ensuring that the learned latent embeddings remain consistent with the underlying causal structure while capturing additional associative patterns and enhancing predictive performance. We conducted experiments on multiple clinical benchmark datasets and demonstrate that this method significantly outperforms traditional causal discovery techniques and standard VGAEs in terms of predictive accuracy and F1-score. Moreover, the resulting latent embeddings provide interpretable insights into both causally relevant factors and complementary associative relationships, enabling clinicians to better understand underlying disease mechanisms and make informed decisions. Overall, our proposed approach combines causal inference and deep graph learning to generate interpretable latent embeddings, improve disease prediction, and support personalized medicine.</p>