Cargando…

A Machine Learning Model for the Accurate Prediction of 1-Year Survival in TAVI Patients: A Retrospective Observational Cohort Study

Background: predicting the 1-year survival of patients undergoing transcatheter aortic valve implantation (TAVI) is indispensable for managing safe early discharge strategies and resource optimization. Methods: Routinely acquired data (134 variables) were used from 629 patients, who underwent transf...

Descripción completa

Detalles Bibliográficos
Autores principales: Pollari, Francesco, Hitzl, Wolfgang, Rottmann, Magnus, Vogt, Ferdinand, Ledwon, Miroslaw, Langhammer, Christian, Eckner, Dennis, Jessl, Jürgen, Bertsch, Thomas, Pauschinger, Matthias, Fischlein, Theodor
Formato: Online Artículo Texto
Lenguaje:English
Publicado: MDPI 2023
Materias:
Acceso en línea:https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10488486/
https://www.ncbi.nlm.nih.gov/pubmed/37685547
http://dx.doi.org/10.3390/jcm12175481
Descripción
Sumario:Background: predicting the 1-year survival of patients undergoing transcatheter aortic valve implantation (TAVI) is indispensable for managing safe early discharge strategies and resource optimization. Methods: Routinely acquired data (134 variables) were used from 629 patients, who underwent transfemoral TAVI from 2012 up to 2018. Support vector machines, neuronal networks, random forests, nearest neighbour and Bayes models were used with new, previously unseen patients to predict 1-year mortality in TAVI patients. A genetic variable selection algorithm identified a set of predictor variables with high predictive power. Results: Univariate analyses revealed 19 variables (clinical, laboratory, echocardiographic, computed tomographic and ECG) that significantly influence 1-year survival. Before applying the reject option, the model performances in terms of negative predictive value (NPV) and positive predictive value (PPV) were similar between all models. After applying the reject option, the random forest model identified a subcohort showing a negative predictive value of 96% (positive predictive value = 92%, accuracy = 96%). Conclusions: Our model can predict the 1-year survival with very high negative and sufficiently high positive predictive value, with very high accuracy. The “reject option” allows a high performance and harmonic integration of machine learning in the clinical decision process.