Identification of Patients with Potential Atrial Fibrillation during Sinus Rhythm Using Isolated P Wave Characteristics from 12-Lead ECGs

Figures

• 

  Figure 1. Patient selection for the AF and control groups.

• 

  Figure 2. Flowchart of ECG processing and feature extraction in this study.

• 

  Figure 3. P wave isolation by applying the MA-UPEMD method. (a) The first step in MA-UPEMD: a sinusoid signal is added on the input signal, resulting in a masked signal. (b) Plots demonstrating one of the sifting iterations in MA-UPEMD. UPEMD is applied with the minimum arclength criterion on the QRS peaks. Specifically, when interpolating the spline for upper and lower envelope, the knots for Q, R, and S peaks (o) are replaced with new knots found by applying the minimum arclength criterion (*). After taking the mean of the upper and lower envelopes, the P and T waves are derived (purple). Subtracting this mean gives the QRS wave (black). This is only one sifting for one phase in MA-UPEMD. The process is repeated four times with different uniform phases to derive the mean signal with only P and T waves (c). Then, the same MA-UPEMD is applied again on the signal with P and T waves, with P wave as the spike to be separated. This results in the delineated baseline wonder and T wave (purple) in (d) and the extracted P wave (red) in (e).

• 

  Figure 4. P wave extraction using the bandpass filter method.

• 

  Figure 5. The PCA transform for eight of the 12 leads. (a) The eight channels for PCA transform (b) The 3 PCs—PC1, PC2, and PC2, respectively. (c) The three-dimensional space constructed by the new eigen space. The trajectory is defined by the three PCs in (b). Note that only the P wave is used when calculating the transformation matrix in PCA. The QRS and T waves in (b,c) are projected by the same matrix constructed by P wave and are only for demonstrating the relative position of the P wave.

• 

  Figure 6. The P-loop descriptor by PC1 and PC2. The number of grey cells is the loop length, while the number of white cells inside the loop is the loop area.

• 

  Figure 7. The inter-lead correlation of the P-loop by PC1, PC2 and PC3. The left panel is an example from a control subject and the right panel is an example from a patient in the AF group.

• 

  Figure 8. Comparison of P wave signals of the same ECG extracted by MA-UPEMD and the bandpass filter method in the time domain (a) and their corresponding time-frequency representations (b–d). The adjacent R wave, which has a wide frequency band as shown in (b), introduces power leakage in the bandpass filter method (d). In contrast, MA-UPEMD separates the R wave at a precise time and frequency range, and thus better preserves the P wave.

• 

  Figure 9. ROC curve and the optimal operating point of machine learning classifiers for AF prediction on the testing dataset. This figure shows the ROCAUC of SVM, random forest (RF), XGBoost (XGB), and perceptron classifiers using (a). features extracted with the proposed method (b). features extracted with the bandpass filter method.

• 

  Figure 10. ROC curve and the optimal operating point of different machine learning classifiers (SVM, random forest (RF), XGBoost (XGB)) with P wave amplitude and duration.

• 

  Figure 11. The top ten features that contribute the most SHAP feature importance in perceptron model. The abbreviations of feature names are explained in Table 1.