Electrocardiogram lead selection for intelligent screening of patients with systolic heart failure

Figures

• 

  Figure 1. Data selection flowchart. A total of 900 patients with systolic HF were included in the study. For comparison, 900 age-matched individuals without systolic HF were included in the research. After ECG preprocessing, noisy data and data with ECG splitting errors were excluded. The remaining 700 patients with systolic HF and the corresponding individuals without HF were combined into one dataset. These data were then separated into groups of 1260 for training data, 140 for validation data, and 186 for testing data.

• 

  Figure 2. Signals with excessive ECG noise, such as R peak detect error, electromyogram noise, baseline wandering and power line artifact, were excluded from our study. These ECG signals are severely affected by noise and therefore cannot be processed to obtain the correct ECG compound. They were excluded from our research to avoid misleading the neural network model during training.

• 

  Figure 3. Illustration of the research process employed in this study. The first step was to extract the ECG signal from the JPG images. The second step was to transform each single-lead ECG signal into CWT spectra. In the final step, the spectra were trained separately in 12 models for the 12-leads, and the softmax layer output scores (ranging from 0 to 1) were recorded and applied for the comprehensive scoring method. Four comprehensive scoring methods were considered, including one where equal weights were given to the 12-leads and the key leads close to the left ventricle (Leads I, V5, and V6).

• 

  Figure 4. 12-lead ECG signal preprocessing. Details are provided in seven steps. Before converting the 12-lead ECG images into the signal, it was necessary to confirm the target area of the ECG signal and then crop the region containing the ECG image and transform the RGB image into a grayscale image. Image binarization could convert the color of the signal and background into black or white. Then, the pixel bits of each lead signal were regarded as coordinates: X, regarded as time, and Y, regarded as relative size, were independently accessed. The squares on the 12-lead ECG images were used as the basis for determining the length of time required to reconstruct the original signal. Finally, the interpolated signal was used to compare the relative position of the pixel to the time required to reconstruct the electrical signal and the signal’s size.

• 

  Figure 5. Illustration for the CWT of the 12-lead ECG. The original image of the ECG JPG file and CWT spectra were showed.

• 

  Figure 6. The ROC curves were presented. (A) presents a comparison of the 12-lead results for each lead and (B) presents the comprehensive scoring method results.