Artificial intelligence to improve decision making in transcatheter aortic valve implantation

Abstract

Abstract Background The current state of the art in the diagnosis and treatment of cardiovascular diseases has been based on evidence resulting from traditional trials as well as years of clinical experience. Due to interindividual differences and a huge number of possible cofounders, in interventional cardiology a linear algorithm is usually not able to precisely estimate individual risk, therapy, or outcome. With the technological evolution in deep machine learning (ML) and artificial intelligence (AI), clinicians may now address aspects that might not have been investigated previously, as supercomputers may handle the plethora of data that are generated as part of treatment. Ultimately, treatment recommendations and decisions may be made on a personalized level. Purpose The aim of this study was to apply AI to routine clinical practice to improve decision making in transcatheter aortic valve implantation (TAVI) to determine the best type and size of prosthesis personalized for each patient with pre-procedural risk stratification. Methods All patients included in the study were undergoing TAVI. To predict the clinical outcomes we applied a random forest classification, a ML method with high interpretability. For baseline data 58 features were chosen, including valve type and size used. After removing highly collinear features with a certain variance inflation factor, feature selection was based on impurity-based feature importance as well as permutation importance. The performance of the estimators was evaluated by a five-fold nested stratified cross-validation. To evaluate the model ROC and mean AUC scores were chosen. Results A total of 3882 patient datasets were included in this trial. The baseline characteristics were consistent with a high cardiovascular risk typical of this collective. Device success was achieved in 83.3%, pacemaker implantation was necessary in 12.2%, and aortic valvular insufficiency was observed in 2.5%. The 30-day mortality was 3.4% and one-year mortality was 12.7%. The mean AUC for the outcome parameters device success, aortic valvular insufficiency, any pacemaker operation, and 30-day and one-year mortality after five-fold cross validation were 0.61±0.03, 0.71±0.04, 0.66±0.04, 0.67±0.03, and 0.69±0.01, respectively. Conclusions We report preliminary data concerning a promising method to improve decision making in the context of TAVI evaluation and planning using ML algorithm. We showed the feasibility with acceptable AUC values for all outcome parameters. Thus, the integration of AI in TAVI strategy planning process offers a valuable tool providing patient focused personalized therapy. Funding Acknowledgement Type of funding sources: None.