Abstract 13957: Predicting In-Hospital Mortality Among Patients Hospitalized With Heart Failure Using a Convolutional Neural Network and Transfer Learning on a Single Admission Chest X-Ray

Background: Early identification of mortality risk among patients hospitalized with heart failure (HF) is critical for therapeutic decisions. Early warning systems have relied on large datasets with frequent outcomes. We evaluate the use of transfer learning to use a single chest x-ray obtained at presentation in a small sample of HF hospitalizations to predict in-hospital mortality. Methods: In MIMIC-IV - a publicly available electronic health record dataset of consecutive hospitalizations from the ED or those requiring ≥1 day of ICU stay during 2008-2019 - we identified all hospitalizations with a primary diagnosis of HF. We combined this with chest x-rays (CXRs) drawn from the MIMIC-CXR dataset with available radiographic labels. We evaluated the risk of mortality based on CXR using a convolutional neural network (CNN) based on the Efficient Net B3 architecture. We compared 2 strategies - (1) training on primary HF hospitalizations directly for mortality risk (direct learning), or (2) pre-training the model on a non-mortality radiographic label, before training for mortality (transfer learning) (Fig A). Results: There were 3604 hospitalizations with a primary diagnosis of HF (mean age 70.8 years, 48% women) who had a corresponding CXR, of whom 87 (2.4%) died in the hospital. There were 102,551 hospitalizations without a diagnosis of HF with a CXR. The CNN model developed on HF data (60% training, 20% validation, 20% test) led to modest predictive performance (AUROC 0.67) for mortality (Fig B). In the second model that was pre-trained on identifying pleural effusion on CXRs in the absence of a HF diagnosis, and was tuned on the HF data for mortality, was associated with a large predictive gain (AUROC 0.80) (Fig B). Conclusion: A single CXR can accurately classify the mortality risk of patients with HF with high model discrimination. Transfer learning allows the use of unstructured data inputs in predictive models for small datasets and few inputs.