Despite saving many people’s lives with heart failure (HF), surgical, mechanical ventricular support, and cardiac replacement therapies include substantial risks. Despite this early, less selective surgical referral has been on the rise due to the challenges of identifying medical therapy nonresponders on time and the severe implications of nonresponse. As a result, patients who might have improved with medicinal treatment are forced to undergo the risk and disruption of surgical treatment. This research aimed to help clinicians quickly and accurately identify HF medical therapy nonresponders by developing deep learning models based on commonly available EHR characteristics. Patients (18-90 years) admitted to a single tertiary care facility between January 2009 and December 2018 with International Classification of Disease HF diagnostic coding made up the study cohort. Standard electronic health record (EHR) data was used to train ensembles of deep learning models using time series and densely connected networks. All observations that resulted in a serious progression (death from any cause or referral for HF surgical intervention) within 1 year were included in the positive class. For model training, validation, and testing, researchers used more than 350 million EHR data points collected from 79,850 unique hospitalizations of 52,265 HF patients who met the observation criteria. About 20% of the observations made by the model meet the positive class requirements. The C-statistic for the model was 0.91. Clinical relevance is supported by the demonstrated accuracy of EHR-based deep learning models in predicting 1-year all-cause death or referral for HF surgical therapy. There is significant promise for EHR-based deep learning models to help HF practitioners better apply advanced HF surgical therapy in nonresponders to medicinal therapy.