Machine learning vs. conventional methods for prediction of 30-day readmission following percutaneous mitral edge-to-edge repair

Abstract

Background: Identifying predictors of readmissions after mitral valve transcatheter edge-to-edge repair (MV-TEER) is essential for risk stratification and optimization of clinical outcomes.

Aims: We investigated the performance of machine learning [ML] algorithms vs. logistic regression in predicting readmissions after MV-TEER.

Methods: We utilized the National-Readmission-Database to identify patients who underwent MV-TEER between 2015 and 2018. The database was randomly split into training (70 %) and testing (30 %) sets. Lasso regression was used to remove non-informative variables and rank informative ones. The top 50 informative predictors were tested using 4 ML models: ML-logistic regression [LR], Naive Bayes [NB], random forest [RF], and artificial neural network [ANN]/For comparison, we used a traditional statistical method (principal component analysis logistic regression PCA-LR).

Results: A total of 9425 index hospitalizations for MV-TEER were included. Overall, the 30-day readmission rate was 14.6 %, and heart failure was the most common cause of readmission (32 %). The readmission cohort had a higher burden of comorbidities (median Elixhauser score 5 vs. 3) and frailty score (3.7 vs. 2.9), longer hospital stays (3 vs. 2 days), and higher rates of non-home discharges (17.4 % vs. 8.5 %). The traditional PCA-LR model yielded a modest predictive value (area under the curve [AUC] 0.615 [0.587-0.644]). Two ML algorithms demonstrated superior performance than the traditional PCA-LR model; ML-LR (AUC 0.692 [0.667-0.717]), and NB (AUC 0.724 [0.700-0.748]). RF (AUC 0.62 [0.592-0.677]) and ANN (0.65 [0.623-0.677]) had modest performance.

Conclusion: Machine learning algorithms may provide a useful tool for predicting readmissions after MV-TEER using administrative databases.

Keywords: Machine learning; National Readmission Database; Transcatheter edge-to-edge repair.