Heart disease is one of the major causes of death worldwide and predictive models that are applicable within a wide range of patients are needed; however, existing machine learning and deep learning approaches rely on shallow network structures or single-source datasets and lack robustness in heterogeneous clinical settings. This paper presents a Hybrid Deep Neural Network (HDNN) which combines convolutional neural networks (CNNs) with hierarchical interaction of feature with long short-term memory (LSTM) networks with gated dependencies together with a relevance-sensitive weighting scheme and risk-sensitive loss to achieve enhanced optimization stability and clinical sensitivity. The framework is evaluated based on the dual-scale validation framework that includes the UCI Cleveland benchmark dataset and a multi-source dataset constructed by integrating of five publicly available heart disease repositories. The proposed HDNN evaluates at 97.75 % (AUC =0.9885) accuracy on Cleveland and 98.86 % (AUC = 0.9978) on the multi-source dataset and, overall, it demonstrates superior performance over the baseline machine and deep learning models with clinically relevant indicators. These results demonstrate that the suggested framework provides a robust and reliable generalizable method of heart disease risk prediction with promising implications for clinical decision support.