Machine learning for enhancing the ablation outcomes using recurrence quantification analysis in human persistent atrial fibrillation

Figures

• 

  Figure 1. The pipeline of the proposed method. (A) Data collection using Ensite array catheter (B) Resampling the data to 512 Hz to reduce the processing time and also the memory allocation (C) QRST subtraction process. (D) Feature extraction using nine RQA variables. (E) Feature normalization using the standardization method. (F) Order the features based on their importance. (G) The feature sets that were used to train and test the model. (H) GridsearchCV technique for hyperparameters tuning (I) Training and testing the 10 ML models (J) Assessment to select the high-performance model. (K) Predicting the local atrial tissue responses to catheter ablation.

• 

  Figure 2. Selection the optimal values for (A) delay time, (B) embedded dimension, and (C) phase space trajectory plot for real EGM from patient 7 with positive response to catheter ablation.

• 

  Figure 3. Selection of the time delay using the AMI. The solid line represents the AMI calculated from the original EGM signal. In contrast, the dashed line represents the mean AMI calculated from the 100 randomly generated surrogate signals generated from the original EGM. The shaded regions indicate the significance of 95% confidence intervals. The surrogate test preserves the amplitude distribution while disrupting the temporal dependencies to estimate the finite-sample bias floor.

• 

  Figure 4. Top-ranking of the RQA features. FS1, FS2, …, FS9 are the feature set 1, 2, …9, respectively.

• 

  Figure 5. Splitting the data into train and test sets using leave EGMs one patient out a 10-fold cross validation (LOPOCV).

• 

  Figure 6. (A) and (B) are the distance matrices and the corresponding RPs for the real EGMs collected from patient 7 with the negative and positive responses to ablation, respectively.

• 

  Figure 7. The RQA measures calculated from the EGM positive and negative responses to catheter ablation. (A) The RR variable, (B) and (C) are the variables related to diagonal and vertical lines, respectively. P, Positive; N, Negative.

• 

  Figure 8. Models’ performance using test data set in the evaluating process.

• 

  Figure 9. (A) the accuracy for each patient; (B) the mean and the standard deviation of the overall accuracy.

• 

  Figure 10. Confusion matrix for the DT model using FS3.

• 

  Figure 11. ROC and the AUC for the DT using FS3.