A novel adaptive CNN-LSTM fusion network for electrocardiogram diagnosis

Objective. Cardiovascular disease (CVD) causes severe global health threat, and electrocardiogram (ECG) is crucial for early CVD diagnosis. Recently, two popular deep learning methods, that is, convolutional neural network (CNN) and long short-term memory (LSTM) network are studied for ECG modeling and CVD diagnosis, but CNN adopts fixed kernels, thereby reducing efficiency and introducing noise, and LSTM struggles with local feature correlations. Approach. This study proposes an adaptive CNN-LSTM (aCNN-LSTM) fusion network for ECG diagnosis. An adaptive convolutional kernel is newly designed, which can dynamically adjust size based on local signal variance. Smaller kernels optimize efficiency in stationary segments, while larger kernels extract diverse features in non-stationary regions. The adaptive features from aCNN are further fed into LSTM to capture temporal relationships. Finally, a spatial-temporal fusion mechanism is used and a multi-class classification is achieved via the output layer. Main results. Experiments on the PTB-XL dataset show that the proposed aCNN-LSTM net outperforms CNN, LSTM, and CNN-LSTM in diagnosis performance: its overall accuracy reaches 89.89%, macro-average F 1-score is 0.9640, and weighted-average F 1-score is 0.9698. Significance. This method enhances the efficiency and accuracy of automatic ECG diagnosis, and provides reliable technical support for early CVD screening in clinical and primary medical settings.