| CPC A61B 5/346 (2021.01) [A61B 5/7267 (2013.01); G16H 50/30 (2018.01)] | 8 Claims |
|
1. A method for predicting multi-type electrocardiogram (ECG) heart rhythms, the method comprising:
(1) acquiring 12-lead ECG signals from a body surface of a patient, and resampling an ECG signal of each lead to a same signal length;
(2) constructing a node mutual information pooling U-shaped graph convolution network, and extracting deep features of the ECG signals by using a feature extraction module in the graph convolution network;
(3) performing one-layer one-dimensional convolution on the deep features to obtain a graph feature matrix to be constructed, where rows of the graph feature matrix correspond to various types of arrhythmias to be predicted, and columns of the graph feature matrix correspond to feature vectors corresponding to the various types of arrhythmias; and, converting the graph feature matrix into an undirected graph G(V, A, E, W, X), where V is a point set of the undirected graph, and each node in the point set corresponds to each type of arrhythmias; E is an edge set of the undirected graph, which records a similarity between nodes; W is a weight matrix, which gives different weights to different edges; X is a node feature matrix; and, A is a node adjacency matrix;
(4) inputting the obtained undirected graph into a graph encoding module in the graph convolution network, quantitatively calculating node mutual information of the undirected graph by using the graph encoding module, and selecting a node subset with maximum mutual information to decrease a number of nodes in the undirected graph for down-sampling;
(5) inputting the undirected graph with decreased number of nodes into a graph decoding module in the graph convolution network, up-sampling, by using the graph decoding module, the undirected graph according to existing node indexes to restore an original number of nodes so as to generate a new graph feature matrix, and adding the new graph feature matrix with the graph feature matrix before down-sampling and outputting an aggregate graph feature matrix;
(6) performing summation, maximization, minimization and averaging on the aggregate graph feature matrix, splicing a result obtained after four operations into a feature vector and inputting the feature vector into a fully-connected layer, and finally mapping to obtain a probability value of each type of arrhythmias; and
(7) training the graph convolution network through a cross entropy loss function by using a large amount of ECG data and labels of various arrhythmias to predict multi-type ECG heart rhythms.
|