Machine learning-based analysis of risk factors for atrial fibrillation recurrence after Cox-Maze IV procedure in patients with atrial fibrillation and chronic valvular disease: A retrospective cohort study with a control group

Figures

• 

  Figure 1. Kaplan–Meier curves were used to visualize the sinus rhythm maintenance rate in the two groups. The Log Rank rank sum test statistic for the comparison of survival time between the two groups was 673.932, indicating a significant difference in sinus rhythm maintenance rate between the CMP-IV group and the Non-CMP-IV group (p < 0.001). The median sinus rhythm maintenance time for the CMP-IV group was 88 months (95% CI: 84, NA) and for the Non-CMP-IV group was 8 months (95% CI: 7, 8). The hazard ratio for the Cox-Maze IV group compared to the Non-Cox-Maze IV group was 0.092 (95% CI: 0.074, 0.114). The follow-up time was plotted in months, and the number of patients at risk was shown below the graph.

• 

  Figure 2. The receiver operating characteristic (ROC) curves of the 10 machine learning models established in the two groups. Panel A shows the ROC curves for the CMP-IV group, while Panel B shows the ROC curves for the No CMP-IV group. The acronyms in the legend stand for the following: SVM, support vector machine; Catboost, categorical boosting; XGBoost, extreme gradient boosting; AdaBoost, adaptive boosting; Bagging, bootstrapped aggregation; GBDT, gradient boosting decision trees; LightGBM, light gradient boosting machine; MLP, multilayer perceptron.

• 

  Figure 3. The feature importance plots of 4 machine learning models from the CMP-IV group are shown in A–D. Each plot is a bar graph with the feature importance on the x-axis and the names of each feature on the y-axis. The x-axis represents the feature importance, with the values indicating the contribution of the feature to the model. A higher value on the x-axis indicates a higher importance of the feature in the respective model. The models used are XGBoost (A), LightGBM (B), CatBoost (C), and Random Forest (D).

• 

  Figure 4. The feature importance plots of 4 machine learning models from the Non-CMP-IV group are shown in A–D. The models used are XGBoost (A), LightGBM (B), CatBoost (C), and Random Forest (D).

• 

  Figure 5. The SHAP force plots shown in A–D correspond to model-predicted risk scores for particular instances. The SHAP force plot is a graphical representation of the feature importance for a single sample in a machine learning model. The x-axis represents the different features, arranged from left to right. Each feature has a corresponding weight value, which indicates the impact of that feature on the prediction outcome. The size of the weight value indicates the magnitude of the impact of the feature on the prediction. The reference line represents the expected value of the prediction outcome if all feature values were set to their average value. The f(x) value represents the predicted outcome for the given sample. A,B show high and low risk examples for the CMP-IV group, respectively. C,D show high and low risk examples for the non-CMP-IV group, respectively.

• 

  Figure 6. SHAP summary plots based on XGBoost model output (A) and catBoost model output (B) from the CMP-IV group. SHAP summary plots are used to visualize the impact of each feature on the model output. The plot includes a list of features on the left side and the corresponding feature values on the right side. The middle part of the plot is a cluster of coloured dots, where the red dots represent features that have a positive impact on the model output, the blue dots represent features that have a negative impact, and the purple dots represent features that have a neutral impact. The size of the dots represents the magnitude of the impact of each feature, with larger dots indicating a greater impact. The influence of each feature on the outcome can be discerned by comparing the position of the dots to the reference line.