Multimodal Machine Learning-Based Technical Failure Prediction in Patients Undergoing Transcatheter Aortic Valve Replacement

Figures

• 

  Central Illustration. Multimodal Machine Learning-Based VARC-3 Technical Failure Prediction in Patients Undergoing TAVR Study population (top), machine-learning workflow (middle), and the best-performing machine-learning models for predicting VARC-3 technical failure (bottom). ROC curves and confusion matrixes show the best-performing models for predicting each category of technical failure: cardiac technical failure (Kruskal-Wallis – Random Forest) and vascular technical failure (Kruskal-Wallis – Random Forest). Angio = angiography; AUC = area under the curve; CT = computed tomography; ECG = electrocardiogram; Echo = echocardiography; ML = machine-learning; ROC = receiver operating characteristic; SHAP = SHapley Additive exPlanations; SMOT = Synthetic Minority Oversampling Technique; STS-PROM = Society of Thoracic Surgeons Predicted Risk of Mortality; TAVR = transcatheter aortic valve replacement; VARC = Valve Academic Research Consortium.

• 

  Figure 1. Heat Maps of Model Performance Heat maps illustrating accuracy, AUC, sensitivity, specificity, and negative predictive value for cardiac (top) and vascular technical failure (bottom) across different classifiers and feature selection methods. Classifiers include AdaBoost (AB), Autoencoder (AE), linear discriminant analysis (LDA), logistic regression (LR), Lasso logistic regression (LRLasso), and random forest (RF). Feature selection methods include analysis of variance (ANOVA), Kruskal-Wallis (KW), recursive feature elimination (RFE), and Relief. AUC = area under the curve; NPV = negative predictive value.

• 

  Figure 2. Predictive Performance of Top 4 Models for VARC-3 Technical Failure ROC curves for the top 4 models predicting VARC-3 technical failure (top) and statistical comparison of AUCs among the models (bottom). Dark blue indicates statistically significant differences (P < 0.05). TAVR = transcatheter aortic valve replacement; VARC = Valve Academic Research Consortium; other abbreviations as in Figure 1.

• 

  Figure 3. SHAP Summary Plot for VARC-3 Technical Failure SHAP summary plots illustrating the impact of each features on the best-performing models for cardiac (left) and vascular (right) technical failure. These plots visualize the contribution of each features to the predictive performance of each model. AVC = aortic valvular complex; CVE = cerebrovascular event; LVOT = left ventricular outflow tract; SHAP = SHapley Additive exPlanations; STS-PROM = Society of Thoracic Surgeons Predicted Risk of Mortality; other abbreviations as in Figure 1, Figure 2.