Predictors of Length-of-Stay Among Transcatheter Aortic Valve Replacement Patients Using a Supervised Machine Learning Algorithm

Abstract

Background

Length of stay following transcatheter aortic valve replacement (TAVR) continues to improve, but significant gaps remain in predicting the length of stay following TAVR.

Objectives

This study aimed to develop a novel machine learning (ML) algorithm that would facilitate the understanding of the predictors of early and late hospital discharge in patients who have undergone TAVR.

Methods

Using the Biome data set, 9,172 outpatient TAVR procedures were analyzed from 21 centers between 2017 and 2021 across the United States. Supervised random forest ML algorithms were developed to identify variables involved in short length of stay (SLOS) (length of stay <36 hours) and long length of stay (LLOS) (length of stay ≥72 hours) in a 70% sample of the Biome data set. The models were then tested on the remaining 30% of the data set and results compared to standard multivariable models in predicting LOS.

Results

Twenty and 22 variables were identified and included as important predictors for the SLOS and LLOS multivariable models, respectively. The predictive power of both the SLOS (sensitivity 0.81, specificity 0.70, area under the curve [AUC] 0.82) and LLOS (sensitivity 0.45, specificity 0.94, AUC 0.85) ML models were more robust than the standard multivariable model (SLOS AUC 0.65, LLOS AUC 0.65). Our study uncovered several previously unreported predictors for length of stay following TAVR, such as procedural duration, postprocedure physical therapy, and procedure day of the week.

Conclusions

ML algorithms may have an important role in identifying novel predictors of short and prolonged length of stay following TAVR. These efforts may facilitate targeted quality improvement programs to decrease length of stay post-TAVR.