基于变权重奇异谱分析的心律不齐识别方法

现有心律不齐研究多数围绕心电信号中不同频率特性成分的分离展开，而不同子序列的信息量对于最终目标决策的贡献则缺少研究与分析.为增强高贡献度子序列对于分类器的影响，提出了一种变权重奇异谱分析与深度学习结合的识别方法.通过奇异谱分析获得多个子序列，结合各个子序列的奇异值计算随机森林下的基尼系数，并将其作为权重.变权重的序列样本用于训练神经网络模型，更高效地挖掘了有用信息，进一步提高了识别精度，最终的心律不齐识别准确率为98.35%，Macro-F1为97.95%.相对于传统的定值权重，本文提出的变权重识别方法在各个性能指标上均有明显提升.

Recognition Method of Arrhythmia Based on Variable Weight Singular Spectrum Analysis

Many existing arrhythmia researches focus on the separation of different frequency characteristic components in the ECG signal. However， the contribution of different subsequences to the final target decision-making is lack of research and analysis. In order to enhance the impact of high-contribution subsequences on the classifier， a recognition method combining variable weight singular spectrum analysis and deep learning is proposed. Multiple subsequences are obtained through singular spectrum analysis.The Gini coefficient under the random forest is calculated by the singular value of each sequence and used as the weight. The sequence samples with variable weights are used to train the neural network model， which can mine useful information more efficiently and further improve the recognition accuracy. The accuracy rate of final arrhythmia recognition is 98.35%， and Macro-F1 is 97.95%. Compared with the traditional fixed weight， the proposed recognition method of variable weight has a significant improvement in various performance indicators.