Heart disease, more commonly known as cardiovascular disease (CVD), is a leading cause of mortality besides disability internationally, heart disease claims the lives of 18 million people annually. They can save lives by finding those at risk of heart disease and giving them the treatment they need before it’s too late. There is an urgent need to develop more precise and efficient detection tools for cardiovascular illnesses since they pose a major threat to public health around the world. Nowadays, medical professionals rely heavily on machine learning algorithms, particularly for disease diagnosis using medical databases. Such powerful algorithms and data processing methods have great promise for the accurate prognosis of cardiovascular disease, among other disorders. Thus, this research builds a deep learning model for heart disease estimates using feature selection as an input. For the feature selection procedure from the pre-processed datasets, a modified version of the freshly introduced crayfish optimization algorithm (COA) is suggested. As a last step, this research introduces a new framework for autoencoder-based heart disease detection, the deep attention-based autoencoder (DA-AE). In order to improve feature learning during decoding, the decoder side of DA-AE employs the global attention method. The results demonstrate that our suggested method outdoes the state-of-the-art in terms of accuracy, precision, and recall, besides the F1-score, and that the projected basis offers substantial benefits in the estimate of heart disease. The experiments are conducted on two publicly available datasets.