Current methods for predicting clinical outcomes in heart failure patients often rely on correlational models, overlooking the potential to leverage causal information in improving the robustness of predictions and interpretability. This study introduces an experimental framework to evaluate causally informed classification strategies for predicting mortality in heart failure patients within 28 days, 3 months, and 6 months after hospital discharge. Two causally informed pipelines were developed. One used Markov Blanket feature selection in a standard Machine Learning framework, while the other constructed Bayesian Networks from causal graphs, both compared to a traditional Machine Learning benchmark. The causal structure was inferred by applying causal discovery algorithms (BOSS, fGES, and PC) to bootstrapped samples and combining the resulting graphs with a voting-based ensemble to capture the underlying data-generating process. Results show that causally informed strategies can improve interpretability and reduce model complexity, while maintaining performance at a comparable level.