A neural network model based on attention pooling and adaptive multi-level feature fusion for arrhythmia automatic detection.

With the rising incidence of cardiovascular disease, timely detection and treatment are critical for patients with arrhythmias, and the electrocardiogram (ECG) remains a vital tool for diagnosing and monitoring heart health. In automated arrhythmia detection, researchers have made significant progress in intra-patient paradigms. However, challenges persist in the inter-patient paradigm, where existing methods often rely on manually extracted features or exhibit inadequate performance in detecting anomalous categories. Against the above challenges, this paper proposes a neural network model based on Attention Pooling (AP) and Adaptive Multilevel Feature Fusion (AMFF) to enhance the performance for automatic detection of abnormal categories in the inter-patient paradigm. Among them, the attentional pooling mechanism enables the model to focus on the features of key channels and spatial locations, effectively reducing the influence of redundant information; to address the problem of ECG signal scale differences, we designed adaptive multilevel feature fusion (AMFF), which uses weighted multilevel features to achieve adaptive feature fusion and can utilize multilevel features at the same time, thus enhancing the feature expression capability of the model. Based on following the AAMI criteria, we evaluated the proposed model using the MIT-BIH arrhythmia database. The results showed that the model achieved an overall accuracy of 99.32% in the intra-patient paradigm and 93.35% in the inter-patient paradigm. For the inter-patient paradigm, the model not only performs well in N-category classification but also achieves good results in the anomaly categories of S, V, and F. This demonstrates a relatively balanced performance.