Abstract 12927: A Deep Learning Algorithmic Approach to Identify Predictors of In-Hospital Outcomes in Takotsubo Cardiomyopathy

Abstract

Introduction: The use of deep learning (DL) through artificial neural networks is increasingly being employed to analyze large, complex datasets in clinical research allowing for new insights. Takotsubo cardiomyopathy (TTS) has an incidence of 12.4% with a reported in-hospital mortality of 5% in the US. While overall risk factors for the condition have been identified, data regarding in-hospital predictors are scant. We sought to apply a DL Artificial Neural Network (ANN) algorithm to identify predictors of in-hospital mortality in patients with TTS. Methods: The National In-patient Sample is the largest, publicly available, repository of in-patient data comprising a 20% stratified sample of all US hospital discharges. We identified all unweighted hospitalizations with a primary diagnosis of TTS from 2016-2019 using International Classification of Diseases, Tenth Revision diagnosis codes. Data regarding demographics, hospital characteristics, known confounders and associated secondary diagnoses were abstracted. These data were then analyzed using a multilayer perceptron ANN where 70% (4,579) of the dataset was used for training and 30% (1,894) for testing. An independent variable importance analysis was also performed. Results: The training and testing models were able to accurately predict outcomes in 98.6% and 98.2% of the data, respectively. ANN identified covariates with high independent variable importance (normalized importance of >70%) include interstitial lung disease, ventricular fibrillation, end-stage renal disease, 3rd degree AV block, sudden cardiac arrest, and acute PE (Model AUC: 0.952). Conclusions: Deep learning-based algorithms utilizing ANNs have the potential to identify predictors of in-hospital mortality with significant accuracy in large volume databases such as the NIS. Among all the covariates that were analyzed, our model was able to identify covariates associated with high importance in predicting outcomes.