Abstract 17133: The Use of Novel Hematological Markers in Predicting Stroke After TAVR Using a Machine Learning Algorithm

Abstract

Introduction: Aortic Stenosis is the most common valvular disorder with a predominance in the elderly. Trans-Aortic Valve Replacement (TAVR) has been an effective procedure with marked improvement in quality of life for patients. The procedure carries a small, yet clinically significant risk of stroke. The use of Neutrophil-Lymphocyte Ratios (NLR) and Platelet-Lymphocyte Ratios (PLR) have been growing as novel markers of systemic inflammation. We investigated the ability of a machine learning algorithm (Light GBM) to predict and weigh these ratios along with other clinical parameters for prediction of stroke after TAVR. Objective: To demonstrate the efficacy of the Supervised Machine Learning algorithm, Light GBM, in identifying important variables to predict stroke after TAVR. Methods: We performed a retrospective analysis of 291 patients who underwent TAVR from 2015-2019 at Montefiore Medical Center. Age (80±8), 50.2% Female, BMI (28.7 ±6.3). Clinical data was collected through our Hospital EMR. NLR and PLR averages were obtained using the mean of baseline (prior to surgery), Immediate Post-operative, and Post-operative Day 1 values. A supervised machine learning algorithm, Light GBM, used decision tree algorithms with both level-wise growth and leaf-wise growth. The algorithm was trained on 80% of the data and internally validated on the other 20%. Results: We identified NLR and PLR as the second and third most important feature of importance (Table 1) Clinical and demographic features of importance included BMI, Age, and Sex. Our model when internally validated yield a Sensitivity of 75.0%, Specificity of 91.5%, Accuracy of 91.5%, and F1 of 0.75. The AUC for the model was 0.84. Conclusions: Using Novel hematological parameters in conjunction with machine learning algorithms have highlighted important variables in predicting stroke after TAVR. Extrapolated, average NLR and PLR can be an inexpensive tool in stratifying patients those patients most at risk.