For many year, cardiovascular diseases (CVDs) have still remained a leading cause of mortality worldwide, necessitating accurate and early prediction methods for effective treatment. In this study, we propose a novel TopoAttn model which is considered as an innovative framework that can integrate topological data analysis (TDA), specifically persistent homology, with deep learning to enhance CVD prediction. Generally, our proposed TopoAttn in this paper begins by extracting topological features from patient medical diagnosis data, representing them as persistent barcodes to capture the intrinsic geometric and topological structures of the data. These features are then combined with real-world clinical test data to form a comprehensive representation. To better effectively learn and leverage different types of features (topological and data-explicit), our proposed TopoAttn model employs a multi-layer perceptron (MLP)-based architecture, which is extensively powered with the self-attention mechanism. The additional attention mechanism can assist to prioritize the most critical attributes. Finally, we extensively assess the effectiveness of our TopoAttn model through extensive experiments on a real-world/large-scale medical record dataset. This dataset has been collected from Thu Duc hospital, Thu Duc city, Vietnam - within period of 2021–2024. The experimental results demonstrate the efficacy of our proposed TopoAttn model in significantly enhancing CVD predicting tasks, in comparison to traditional machine learning (ML) and deep learning (DL) methodologies.