Department of Systems and Computer Networks, Faculty of Information and Communication Technology, Wrocław University of Science and Technology, Wrocław, Poland.
PLoS One. 2021 Dec 16;16(12):e0260764. doi: 10.1371/journal.pone.0260764. eCollection 2021.
Feature extraction is an important part of data processing that provides a basis for more complicated tasks such as classification or clustering. Recently many approaches for signal feature extraction were created. However, plenty of proposed methods are based on convolutional neural networks. This class of models requires a high amount of computational power to train and deploy and large dataset. Our work introduces a novel feature extraction method that uses wavelet transform to provide additional information in the Independent Component Analysis mixing matrix. The goal of our work is to combine good performance with a low inference cost. We used the task of Electrocardiography (ECG) heartbeat classification to evaluate the usefulness of the proposed approach. Experiments were carried out with an MIT-BIH database with four target classes (Normal, Vestibular ectopic beats, Ventricular ectopic beats, and Fusion strikes). Several base wavelet functions with different classifiers were used in experiments. Best was selected with 5-fold cross-validation and Wilcoxon test with significance level 0.05. With the proposed method for feature extraction and multi-layer perceptron classifier, we obtained 95.81% BAC-score. Compared to other literature methods, our approach was better than most feature extraction methods except for convolutional neural networks. Further analysis indicates that our method performance is close to convolutional neural networks for classes with a limited number of learning examples. We also analyze the number of required operations at test time and argue that our method enables easy deployment in environments with limited computing power.
特征提取是数据处理的重要组成部分，为分类或聚类等更复杂的任务提供了基础。最近，许多信号特征提取方法被提出。然而，许多提出的方法都基于卷积神经网络。这类模型需要大量的计算能力来训练和部署，并且需要大量的数据。我们的工作引入了一种新的特征提取方法，该方法使用小波变换在独立成分分析混合矩阵中提供附加信息。我们的工作目标是将良好的性能与低推理成本相结合。我们使用心电图（ECG）心跳分类任务来评估所提出方法的有效性。实验使用 MIT-BIH 数据库进行，该数据库具有四个目标类别（正常、前庭异位搏动、室性异位搏动和融合搏动）。实验中使用了不同分类器的几种基本小波函数。最佳的方法是通过 5 倍交叉验证和具有显著性水平 0.05 的 Wilcoxon 检验进行选择。使用所提出的特征提取方法和多层感知机分类器，我们获得了 95.81% BAC 评分。与其他文献方法相比，我们的方法优于大多数特征提取方法，除了卷积神经网络。进一步的分析表明，对于学习样本数量有限的类别，我们的方法的性能接近卷积神经网络。我们还分析了测试时所需的操作数量，并认为我们的方法能够在计算能力有限的环境中轻松部署。
