Prediction of Atrial Fibrillation Recurrence after Thoracoscopic Surgical Ablation Using Machine Learning Techniques

Figures

• 

  Figure 1. Schematic representation of nested cross-validation methodology. Initially, the missing data is removed and an iterative imputation is performed in a stratified 5-fold CV (external) using only the training set. The imputation model is further used to imput the test set. After that, an internal CV is performed for the LASSO feature selection and hyperparameter optimization. As the last step, the model is trained with the training set and validated with the test set. An average AUC is reported.

• 

  Figure 2. Average 5-fold ROC of testing on Outcome 1 or Outcome 2 with all or a selection of variables for two years without AF recurrence. ROC receiver operating characteristic curves, AUC area under the curve.

• 

  Figure 3. Average feature importance using SHAP values over the test folds. The amplitude of the SHAP value indicates the importance of the feature for the prediction (negative values means good outcome, positive values means bad outcome). The colours represent the value of the features, with red for high values (or true for binary) and blue for low values (or false for binary). The importance was calculated for Outcome 1 (A,B) and Outcome 2 (C,D) for the best performing models: LR (A,C) and SVM (B,D).

• 

  Figure 4. Subgroup analysis with pre-defined groups based on Outcome 1 and Outcome 2 best prediction models.