Severe cardiovascular diseases can rapidly lead to death. At present, most studies in the deep learning field using electrocardiogram (ECG) are performed on intra-patient experiments for the classification of coronary artery disease (CAD), myocardial infarction, and congestive heart failure (CHF). By contrast, actual conditions are inter-patient experiments. In this study, we proposed a deep learning network, namely, CResFormer, with dual feature extraction to improve accuracy in classifying such diseases. First, fixed segmentation of dual-lead ECG signals without preprocessing was used as input data. Second, one-dimensional convolutional layers performed moderate dimensionality reduction to accommodate subsequent feature extraction. Then, ResNet residual network block layers and transformer encoder layers sequentially performed feature extraction to obtain key associated abstract features. Finally, the Softmax function was used for classifications. Notably, the focal loss function is used when dealing with unbalanced datasets. The average accuracy, sensitivity, positive predictive value, and specificity of four classifications of severe cardiovascular diseases are 99.84%, 99.68%, 99.71%, and 99.90% in intra-patient experiments, respectively, and 97.48%, 93.54%, 96.30%, and 97.89% in inter-patient experiments, respectively. In addition, the model performs well in unbalanced datasets and shows good noise robustness. Therefore, the model has great application potential in diagnosing CAD, MI, and CHF in the actual clinical environment.

Although deep learning methods have recently attracted considerable attention in the medical field, analyzing large-scale electronic health record data is still a difficult task. In particular, the accurate recognition of heart failure is a key technology for doctors to make reasonable treatment decisions. This study uses data from the Medical Information Mart for Intensive Care database. Compared with structured data, unstructured data contain abundant patient information. However, this type of data has unsatisfactory characteristics, e.g., many colloquial vocabularies and sparse content. To solve these problems, we propose the KTI-RNN model for unstructured data recognition. The proposed model overcomes sparse content and obtains good classification results. The term frequency-inverse word frequency (TF-IWF) model is used to extract the keyword set. The latent dirichlet allocation (LDA) model is adopted to extract the topic word set. These models enable the expansion of the medical record text content. Finally, we embed the global attention mechanism and gating mechanism between the bidirectional recurrent neural network (BiRNN) model and the output layer. We call it gated-attention-BiRNN (GA-BiRNN) and use it to identify heart failure from extensive medical texts. Results show that the $F\mathsf{}\mathrm{1}$ score of the proposed KTI-RNN model is 85.57%, and the accuracy rate of the proposed KTI-RNN model is 85.59%.