An artificial intelligence-enabled ECG algorithm for identifying ventricular premature contraction during sinus rhythm

Figures

• 

  Fig. 1. The ECG image processing process before input. a The standard 12-lead ECG image. b The red-rid background of the 12-lead ECG images was removed. c The image was cropped to be focused on the ECG signals

• 

  Fig. 2. The ECG data input format. a The red-grid background of the 12-lead ECG image was removed and ECG was converted to a gray-scale image. b The pixel intensity was inversed and the pixel intensity was made to 255 pixels. The image was cut vertically into four sub-images according to the “start” and “end” position of each lead. c Pixel-wise scanning sub-images and recording the position where the pixel intensity was equal to 255 pixels. d The closest position of the signal was grouped. Each column was split into four values and all values of the columns were synthesized into four lists in each lead. The signals were transferred to be the time-series formats. e The column of each sub-image consisted of 250 pixels. After pixel-wise scanning, one lead with 250 time-series data was formatted. The interpolation operations were used to perform up-sampling for the time-series data (500 Hz, 2.5 s). f The IIR low-pass filter was used to filtered the noise (cut-off frequency = 15 Hz, order = 3). g The magnitude of each lead was normalized into a unified scale

• 

  Fig. 3. The architecture of the convolutional neural network (CNN). a The 2-dimensional image data were processed by five network computer architectures, which included VGG16, ResNet50V2, InceptionV3, InceptionResNetV2, and Xception to get the best image recognition at the Image Net part of CNN and then flattened into Global Average Pooling (GAP). b The dense layer of the single input from the 2-dimensional image data was connected. The dropout was added to avoid overfitting (drop rate = 0.5) and another dense layer with size of two was added to get an output layer. VPC ventricular premature complex, NOR normal rhythm. c The signals of time-series data were extracted through the CNN layers and flattened by GAP. The output features of the single -input model was directly connected to dropout (dropout rate = 0.5) and the multiple-input model from the twelve channels’ features were merged (to get the output result (dense size = 2). GAP Global Average Pooling

• 

  Fig. 4. The accuracy of different image-input CNN models with various sizes of connected layers. CNN convolutional neural network

• 

  Fig. 5. The AUC of the ROC for the combination of the InceptionV3 CNN model and the dense layer (size = 512). CNN convolutional neural network, AUC area under the receiver operating characteristic curve, ROC receiver operating characteristic curve