A hybrid model combining 1D-CNN and BERT for intelligent ECG arrhythmia classification

Arrhythmia is a common cardiovascular disease, whose early diagnosis is crucial to prevent severe cardiac events. Traditional electrocardiogram (ECG) interpretation methods rely on manual analysis, which often suffers from low efficiency and limited accuracy. To address these issues, intelligent algorithms are increasingly being used for automatic arrhythmia recognition. However, many existing methods still face challenges in achieving accurate classification. In this paper, we propose a novel approach that integrates a one-dimensional convolutional neural network (1D-CNN) with Bidirectional Encoder Representations from Transformers (BERT) for arrhythmia classification. The proposed model, named ECGBert, leverages the local feature extraction capability of CNN and the global context modeling strength of BERT. The model enables the precise classification of different types of arrhythmias by performing signal preprocessing, segment encoding, and sequential feature extraction. Experimental results in the MIT-BIH Arrhythmia Database demonstrate that ECGBert significantly outperforms traditional methods and existing hybrid architectures in multiple evaluation metrics. The model effectively captures long-range dependencies between abnormal heartbeats by incorporating the Transformer mechanism. It also maintains an end-to-end learning structure without the need for hand-crafted features, offering strong generalization ability and robustness. This work provides a new methodological framework for intelligent ECG analysis and promotes the innovative application of deep learning in medical signal processing.