Machine Learning for Predicting Mortality in Transcatheter Aortic Valve Implantation: An Inter-Center Cross Validation Study

Figures

• 

  Figure A1. Distribution of missing values per patient on each center population.

• 

  Figure A2. Normalized bar plots with distributions of the categorical features of the two different centers. Non-survived and survived patient distributions are shown in orange and blue, respectively. For all features, missing values are reported in grey, numerical values are identified with pure orange and pure blue, while categorical instances are represented with graduated levels of orange and blue.

• 

  Figure 1. Illustration of the three phases of Transcatheter Aortic Valve Implantation intervention. Originally the catheter crosses the native valve trough transfemoral access (a), follows the deployment and expansion of the valve (b) and finally the expanded valve replaces the native valve (c).

• 

  Figure 2. Diagram of the inter-center cross-validation. The entire exchange process is obtained by organizing a cross-validation protocol in three main stages, which requires the exchange of the models and their evaluation. Each model is validated on the initial center and on the validation center. Finally, a comparison between the internal validation and the external validation results is performed to evaluate the performance of each model on an unseen external center dataset.

• 

  Figure 3. Detailed diagram of the inter-center cross-validation. Training, validation and evaluation are shown in yellow color, while data harmonization, imputation, pre-processing and class-balancing strategy steps are shown in grey color. The inter-center cross-validation protocol is organized in three stages: internal evaluation, external evaluation and final evaluation. The first two stages share a common clinical data processing, a common validation order and evaluation metrics. The internal evaluation stage (left) and the external evaluation stage (right) share a common processing chain including imputation and pre-processing steps. The final evaluation stage (bottom) is the last stage which is necessary to compare the internal validation results with the external validation results. At the first two stages, each model is evaluated, while at the final stage, both internal and external validation results are compared for a final performance assessment.

• 

  Figure 4. Boxplot distributions of the numerical features for the two different centers. AMC and CZE-TU/e feature distributions are shown in blue and orange, respectively. Imputed data for missing values are not included to avoid influence on the original data distribution.

• 

  Figure 5. Area under curve (AUC) of the receiver operating characteristic (ROC) for the top-three classifiers per center with the highest validation score are shown with different colors. Both centers internal validation (a) and external validation (b) AUC metrics are shown jointly in this figure to highlight the similarity.

• 

  Figure 6. Boxplot results considering (a) CZE-TU/e as initial center and AMC as validation center and considering (b) AMC as initial center and CZE-TU/e as validation center. Both the internal and external validations per classifier are shown in blue and orange, respectively.