Abstract 9832: Automated Electrocardiographic Detection of Pulmonary Hypertension Using Deep Learning

作者：Mandar A. Aras, Sean Abreau, Hunter Mills, Lakshmi Radhakrishnan, Liviu Klein, Neha Mantri, Benjamin Rubin, Joshua Barrios, Christel Chehoud, E. I. Kogan, Xavier Gitton, Anderson N. Nnewihe, Deborah A. Quinn, Charles R. Bridges, Atul J. Butte, Jeffrey E. Olgin, Geoffrey H. Tison · 发表于：Circulation · 年份：2021 · DOI：10.1161/circ.144.suppl_1.9832 · 研究领域：Non-Invasive Vital Sign Monitoring

Introduction: Pulmonary hypertension (PH) is a progressive, life-threatening disease, often diagnosed late in its course. Treatments are available for some subtypes, including pulmonary arterial hypertension (PAH). Hypothesis: Deep learning applied to interpretation of electrocardiograms (ECGs) can detect PH and clinically important subtypes. Methods: Adults with either right heart catheterization (RHC) or an echocardiogram (echo) within 90 days before or after an ECG at the University of California, San Francisco from 2012 to 2019 were retrospectively identified and defined as PH (mean pulmonary artery pressure [mPAP] >20 mmHg) or non-PH (mPAP ≤20 mmHg) by RHC or by echo, if RHC unavailable (peak tricuspid regurgitation velocity >3.4 m/s or ≤2.8 m/s). A convolutional neural network (CNN) to detect PH was developed and tested using patients’ 12-lead ECG voltage data. Patients were divided into training, validation, and test sets in a ratio of 7:1:2. The ability of the CNN to detect pre-capillary PH (defined by RHC), PAH (defined as presence of any PAH-specific medication over the 3 months prior and 6 months following mPAP >20 mmHg by RHC), and Group 3 PH (identified by Group 3-consistent International Classification of Disease codes in the 3 months before or after RHC) was also tested. CNN performance was assessed by calculating the area under the receiver operating characteristic curve (AUC), sensitivity, specificity and positive and negative predictive values. Resu...