Cardiovascular Disease (CVD) is a major cause of death worldwide, making it essential to accurately predict CVD in advance for more effective prevention and treatment. Conventional approaches to risk stratification for CVD have limited accuracy, efficiency, and high-dimensional data processing capabilities. However, in recent years, Deep Learning (DL) has emerged as a powerful approach to address some of these challenges, enabling new possibilities for the analysis of heterogeneous datasets, automation of feature extraction, and representation of intricate relationships. This paper details the state-of-the-art developments using DL methods for CVD prediction and examines various architectures, datasets, and evaluation criteria. This paper presents state-of-the-art DL methods for CVD prediction, emphasizing before12 deep neural network architectures like CNNs, RNNs and hybrid models, showing significant improvement in accuracy over traditional methods (DNN accuracy: 84.30% and CNN-LSTM accuracy: 86%). These developments highlight the power of DL to address problems, such as imbalanced data, interpretability, and ways to take advantage of multimodal data. We address the strengths and limitations of state-of-the-art DL solutions and discuss local challenges such as data imbalance, interpretability, and large annotated training datasets. In addition, the paper addresses the future directions of DL in CVD prediction, such as the integration of multimodal data, improved explaining ability of models, and real-time implementation in clinical practice. Through the integration of recent studies, this review has produced key perspectives to facilitate deep learning-based CVD prediction, which can lead to improved early diagnosis, personalized treatment, and patient benefits.