Abstract: The performance of ECG arrhythmia classification mainly depends on both the effective feature extraction and the optimal design of the classifier. Most of the classic methods extract the time domain features or frequency domain features directly to achieve the arrhythmia classification, but the classification performance still needs to be improved for multi-classification tasks. For this issue, the cyclic spectrum analysis technique is used to achieve the multi-arrhythmia classification. The method assumes that the signal is in non-stationary state, and arrhythmia classification can be implemented through establishing a model to capture the hidden period in the ECG signal, which is more appropriate with the actual state of ECG signals. In order to implement the arrhythmia classification, the morphological features and wavelet coefficients time-frequency domain features are extracted. In addition, the cyclic spectrum technology is adopted for extracting the spectral correlation features for the subsequent multi-classification task. Besides, a comparison on the classification performance is also conducted among the artificial neural networks, the traditional support vector machine classifier and extreme learning machine. Experimental results show that the proposed method based on the extreme learning machine can classify ten types of arrhythmias and achieve an average classification accuracy of 98.13% on the MIT-BIH arrhythmia benchmark dataset.

Key words: arrhythmia classification, cyclic spectral, extreme learning machine

摘要： 心电信号心律失常分类性能主要取决于有效的特征提取和分类器设计。针对传统心律失常分类研究中，多数研究直接利用时域或者频域特征实现心律失常分类，对于多类别的分类性能仍有待提高。鉴于此，选用循环谱分析方法实现心律失常多分类任务。假设信号处于非平稳状态，建立更符合心电信号实际状态的模型去捕捉心电信号中的隐含周期实现心律失常分类。在提取形态特征和时频域小波系数特征之外，利用循环谱技术提取了谱相关系数特征用于后续多分类任务。除此之外，比较了人工神经网络、传统支持向量机和超限学习机分类器在该实验环境下的分类性能，通过多组对比实验，结果表明，利用循环谱技术结合超限学习机分类器进行心律失常分类，可以区分10类心律失常并在MIT-BIH心律失常数据库上实现了98.13%的平均分类准确率。

关键词: 心律失常分类, 循环谱, 超限学习机