Early and accurate diagnostic techniques are necessary for cardiovascular diseases (CVDs) to reduce adverse outcomes that remain a significant contributor which led to morbidity and mortality rates, globally. However, the characteristics of medical datasets can lead to misclassifications and lower generalizability, with imbalanced class distributions that affect the performance of predictive models. This study aims to provide a distinct methodology for CVDs using a deep learning model Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM) networks with the Synthetic Minority Oversampling Technique (SMOTE) to further CVD predictive processes. SMOTE allows the creation of synthetic samples, thereby addressing the class imbalance and allowing for an unbiased representation of classes in the training data. MLP is used to learn intricate patterns from structured clinical data, while LSTM captures the temporal dependencies of the data to improve predictive capabilities. The model is trained and assessed on a target CVD dataset with a commendable accuracy of 97.88%. This study evaluated the evidence presented in this combined MLP, LSTM and SMOTE model, which could produce a robust generalizable CVD predictive model.