Detection of Myocardial Infarction Using ECG and Multi-Scale Feature Concatenate

Figures

• 

  Figure 1. The schematic diagram of healthcare application with integration of Internet of Things (IoT) technology. 12-lead electrocardiogram (ECG) information can combine with other physiological signal values from different wearable devices, such as blood glucose, body temperature, and blood pressure, for algorithm models. Physicians can investigate or diagnosis and undertake further action, such as calling an ambulance or ward, through offline monitoring and/or cloud computing.

• 

  Figure 2. Block diagram of the research, including Physikalisch-Technische Bundesanstalt (PTB), 12-lead ECG database gathering, signal preprocessing, data rearrangement, study model networks training, and performance, compared with other studies. The final objective was to build models capable of detecting and/or locating the occurrence of MI by analyzing the 12-lead ECG.

• 

  Figure 3. Block diagram of the proposed method, including signal preprocessing, dataset rearrangement, and model development.

• 

  Figure 4. Samples of healthy control (HC) and inferior myocardial infarction (IMI) 12-lead ECG. IMI can be detected by the abnormalities waveform features of the ECG, as indicated by the arrows that include the ST displacement, T wave inversion, silent Q wave, and so on.

• 

  Figure 5. Multi-lead features-concatenate network. Twelve (12) lead branch convolutional neural networks (CNNs) were used to extract features in each lead independently and were then concatenated and classified.

• 

  Figure 6. Multi-scale features-concatenate network in every lead branch. K represents kernel size, S represents strides. The N-Net only contained structures illustrated in scale 1, and MSN-Net that used two scales will contain the structure of both scale 1 and scale 2. MSN-Net that used three scales contained structures of scale 1, scale 2 and scale 3, and so on.

• 

  Figure 7. Accuracy trends of the proposed myocardial infarction (MI) detection models plotted in shaded error bar. Blue- and gray-shaded areas indicate the range of one standard deviation, and the best average accuracy under the same scale number is marked by an arrow: Accuracy of models using (a) single-scale features; (b) two-scale features; (c) three-scale features, (d) four-scale feature, and (e) five-scale features.

• 

  Figure 8. Accuracy trends of our proposed MI locating models plotted in shaded error bar. Blue- and gray-shaded areas indicate the range of one standard deviation and best average accuracy under the same scale number are marked by an arrow: Accuracy of models using (a) single-scale features, (b) two-scale features, (c) three-scale features, (d) four-scale features, and (e) five-scale features.

• 

  Figure 9. Box plot and significance of the accuracy of the proposed network where 1S indicates models using single-scale features, 9F indicates models using nine filters, and so on. (a) Evidence that supports accuracy decrease due to overfitting caused by excessive model strength. The average accuracy decreased from 95.76% to 94.95% (p < 0.05); (b) Evidence that supports using multi-scale features will increase MI locating accuracy. The average accuracy increased from 60.49% to 61.82% (p < 0.05).

• 

  Figure 10. The performances of the proposed networks were recorded during each epoch where 1S indicates networks using single-scale features, 9F indicates networks using nine filters, and so on. (a) Trained and validated loss of MI detection model during each epoch; (b) Trained and validated the accuracy of MI detection model during each epoch; (c) Trained and validated loss of MI locating model during each epoch; (d) Trained and validated the accuracy of MI locating model during each epoch.

• 

  Figure 11. Comparison between the proposed model and literature where 1S indicates models using single-scale features, 9F indicates models using nine filters, and so on. Single star interprets p < 0.05, double star interprets p < 0.01, and triple star interprets p < 0.001. (a) MI detection accuracy comparison; (b) MI locating accuracy comparison; (c) MI detection sensitivity comparison; (d) MI detection specificity comparison; (e) MI detection F1-score comparison; and (f) MI detection area under the receiver operating characteristic curve comparison.