<p>Heart disease is the fatal cause of non-communicable and often silent deaths worldwide. Early and accurate diagnosis of heart disease is vital to prevent further harm and save patients’ lives. Therefore, a strategy capable of accurately predicting and preventing heart disease before it becomes critical is needed. With the development of artificial intelligence neural networks, their autonomous learning, unsupervised feature modeling, and generalization capabilities can meet complex medical scenarios and provide decision-making assistance. Researchers are encouraged by neural networks in classifying and predicting heart disease. We used the most prevalent and representative Cleveland dataset in our study on neural network prediction models for heart disease. Specifically, the data features were mined and analyzed, and a heart disease prediction model of hybrid neural network based on CNN–BiLSTM–Attention (CBA) was designed and proposed, wherein CNN helps to mine local information of patient digital features, BiLSTM establishes bi-directional communication between local information and the attention mechanism captures the representation vectors of patients’ mutual information for modeling whether they are sick. To evaluate the effectiveness and predictive performance of the developed CBA, we constructed eight ML models as baselines for comparison. Nine metrics are used for longitudinal comparative analysis to demonstrate the advantage of the model in predicting heart disease classification, such as the ROC curve, AUC, precision, recall, F1 score, sensitivity, specificity, accuracy, and MCC. CBA achieved a precision of 96.30%, which is higher than all baseline models. All experiments consistently prove that mining digital features based on CBA significantly ameliorate heart disease prediction performance.</p>