Predicting mortality risk requires clinicians to infer upper demographics and a set of factors linked by comorbidities, sub-consequences through intermediate biomarkers, and potential mechanisms such as frailty or chronic inflammation. Recent advancement of deep learning facilitates to develop powerful prediction models, but most of them ignore this hierarchy to learn latent representations of the data without considering these factors in explicit and clinically interpretable hierarchies. To address this issue, this paper proposes a causally explainable prediction method that leverages hierarchical Bayesian networks with latent variables learned through deep learning. A Transformer-based model with a sparse autoencoder first derives low-dimensional representations to capture clinical patterns, which are used as intermediate nodes between demographic and biomarker levels within a hierarchical Bayesian network. In this way, both latent representations and clinical variables are organized to allow structured and causally informed explanations of mortality risk. Applied to the National Health and Nutrition Examination Survey (NHANES), it recovers risk factors and pathways consistent with biomedical literature and expert-elicited causal diagrams, achieving more than 93% edge-level agreement and clarifying interactions such as sex–albumin effects through clinically plausible demographic mediation.