Attention-based multi-scale features fusion for unobtrusive atrial fibrillation detection using ballistocardiogram signal

Figures

• 

  Fig. 1. Modular framework of the proposed method for AF detection by means of BCG signal. First, the AF and NAF signals were split into the 1-s-segment dataset and the 24-s-segment dataset, respectively. Second, the 1-D morphology feature extracted from Bi-LSTM and the 2-D rhythms feature extracted from PSR were integrated by CNN. Ultimately, the attention mechanism was involved to improve the performance of the AF classification

• 

  Fig. 2. AF and NAF BCG signal: 1-s-segment and 24-s-segment are illustrated. The left part in this waveform is labeled as AF period, the right part is labeled as NAF period, and all periods were split manually by medical experts in accordance to the synchronous ECG signal as a reference. In Fig. 2, the 24-s-segment is divided from the original BCG signal, and the 1-s-segment is extracted from the 24-s-segment by means of taking 0.5 s before and 0.5 s after the occurrence of the J-peak in BCG waveform

• 

  Fig. 3. The ROC curves of the proposed method and the superior ML models in previous method [6] and [9]. The red curve represents the classification performance of the proposed method, the blue curve represents the classification performance of the previous method [6] with the superior RF classifier, and the green curve represents the classification performance of the previous method [9] with the superior RF classifier

• 

  Fig. 4. Structure of the Bi-LSTM network, which is applied to extract 1-D morphology feature. Both Bi-LSTM layers comprised 50 hidden units, to match the 1-s-segment input. The first Bi-LSTM layer reads the data fed from the input layer and outputs a complete sequence at each time step. The second one is configured to only output at the last time step of the sequence, which can be treated as a feature vector for classification. An additional dropout layer with a probability of 0.5 is inserted between the layers to avoid overfitting, ensuring the trained model’s ability of generalization. The output feature vector from the second Bi-LSTM layer is fed into a fully connected layer, which maps the features from 50 dimensions to 2 dimensions. A Softmax function is then employed to the 2-D vector and the followed classification layer gives an AF or NAF result. The initial learning rate is set as 0.0005 and the network training takes 30 epochs

• 

  Fig. 5. Phase space reconstruction of AF and NAF. a, c are the 3-D trajectory of AF and NAF, respectively; b and d are the 2-D projecting plant of AF and NAF, respectively. The 2-D projection is denoted as the 2-D rhythm feature, which represents the different variation of the intervals between AF and NAF. And the overall 2-D projection plot will be used as the input of the deep neural network

• 

  Fig. 6. Integrated framework based on CNN and Bi-LSTM. The 1-D morphology feature extracted from Bi-LSTM network by means of 1-s-segment BCG and the 2-D rhythm feature extracted from PSR by means of 24-s-segment BCG are combined by using the CNN framework. In addition, the attention mechanism will be involved to improve the AF classification performance

• 

  Fig. 7. Structure of the SAM layer. The MATMUL function is calculating the matrix multiplication, the ReLU function is used as the activation function and the Softmax function is utilized for training and normalization. The details of the calculation steps are listed in Formula (10)–(15)

• 

  Fig. 8. Structure of the channel attention module. The stochastic pool function is supplemented to the general average pool and maximum pool functions together, which effectively improves the presentation capability of the network. And the multilayer perceptron (MLP) is utilized to train the network

• 

  Fig. 9. Structure of the spatial attention module. The stochastic pool function is also added to the average pool and the maximum pool functions, and then the three corresponding outputs are spliced together with the same dimension. Finally, the convolutional operation with a 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document}×7 filter and the sigmoid function are utilized to obtain the screening features