Heart disease is a global concern and it is important to address the issue. Early detection and intervention can support an individual's chance of survival. Accurate prediction models are necessary to identify those most at risk. These models support healthcare professionals in providing diagnosing and treatment strategies. This proposed study presents a promising method for heart disease prediction that integrates autoencoders, dropout regularization, neural network clustering, and optimal tuning. The model was evaluated on 10,000 patient records and achieved 92.1% accuracy. The proposed study first groups the patients in the dataset into separate groups based on important features. The results show different patient phenotypes that are used for understanding different types of heart disease. The cluster labels are further encoded as features and then added to the dataset. The following step is to use autoencoders to learn the underlying data representation and recompile the input data. Meaningful features from the data can be extracted. To prevent overfitting, dropout regularization is used. The model is more likely to generalize to the new data and not the specific features of the training data. Next, a genetic algorithm is used to identifying the optimal hyperparameters for the neural network model. This confirms if the model is optimized for performance and generalizability. The conclusions of the study suggest that the proposed approach is a promising technique for heart disease prediction. Out future work should focus on evaluating the model on larger datasets and exploring the use of the model for early detection and intervention.