Pacing Electrocardiogram Detection With Memory-Based Autoencoder and Metric Learning

Figures

• 

  Figure 1. The ECG abnormality detection via the proposed MAE. A sample of training on the pacing ECG database, the memory module records the latent features of the corresponding training ECG type. Given an input sample of other types, the MAE queries the most correlation latent features in the memory module for data reconstruction, resulting in an output significantly different from the input data. But if the input sample belongs to the training type. The MAE can reconstruct very well according to the memory module.

• 

  Figure 2. Illustration of the proposed MAE model. The memory module records latent features and the memory query can obtain the weight of the degree of similarity between the features of input data and the record features. Note that the output of the memory module is the input of the decoder.

• 

  Figure 3. Illustration shows the details architectures of encoder fe and decoder fd in the MAE model.

• 

  Figure 4. Examples ECG (from Lead II) in pacing ECG database. (A) An example shows the morphology of the pacing ECG. (B) An example shows that the morphology of pacing ECG receives other abnormal interference. (C) An example shows that the ECG morphology of the complete left bundle branch block is similar to pacing ECG. Note that sample (A) has no obvious pacemaker nail in lead II, but the proposed MAE method can distinguish the (C) and pacing type by other lead information.

• 

  Figure 5. Visualization of the reconstruction results of AE and MAE on Pacing ECG. (A) The example model is trained on a type of pacing ECG. The input data is a routine ECG. Due to the periodicity of the ECG, the reconstruction data from the traditional AE model is also close to the input data. It is a failure case of ECG reconstruction in terms of error. But the MAE retrieves the training types memory items for reconstruction data and obtains the results significantly different from the input data types. The significant error between the reconstruction data and the input data means that the input data does not belong to pacing ECG. (B) The example model is trained on a type of routine ECG. The input data is a pacing ECG. The traditional AE model gets a failure case on the input data. But the MAE model obtains a higher significant error between reconstructions data and the input data. Therefore, it shows that the input data is a different type from the type of model.

• 

  Figure 6. Scores of objective function obtain by MAE. The score changes immediately when some different types ECG samples in the test data.

• 

  Figure 7. Example ECG signals from the MIT-BIH database. There are five heartbeat classes named N, S, V, F, and Q. The meaning of each letter is shown in Table 4.

• 

  Figure 8. Confusion matrix of MAE model in the MIT-BIH Arrhythmia database. The row labels represent the true class records in each row and the column labels represent the class records predicted by our method. Numbers in each grid show the number of records classified as column labels when its true class is indicated by row label.

• 

  Figure 9. Examples of the MIT-BIH Arrhythmia database. The “N” type and “Q” type are set to the input samples. And the different memory items mean that the training phase records corresponding types of features. The MAE model can get a significant reconstruction error when the input data is different from the recorded data type.