Automated Deep Learning Pipeline for Characterizing Left Ventricular Diastolic Function.

Introduction: Left ventricular diastolic dysfunction (LVDD) is most commonly evaluated by echocardiography. However, without a sole identifying metric, LVDD is assessed by a diagnostic algorithm relying on secondary characteristics that is laborious and has potential for interobserver variability. Methods: To characterize concordance in clinical evaluations of LVDD, we evaluated historical echocardiogram studies at two academic medical centers for variability between clinician text reports and assessment by 2016 American Society of Echocardiography (ASE) guidelines. We then developed a workflow of 8 artificial intelligence (AI) models trained on over 155,000 studies to automate assessment of LVDD. Model performance was evaluated on temporally distinct held-out test sets from two academic medical centers. Results: In a validation cohort of 955 studies from Cedars-Sinai Medical Center, our AI workflow demonstrated 76.5% agreement and a weighted Cohen's kappa of 0.52 with ASE guideline assessment using human measurements. In contrast, the clinician report evaluation had 48.5% agreement and a weighted Cohen's kappa of 0.29 with ASE guidelines. In the Stanford Healthcare cohort of 1,572 studies, the AI workflow had 66.7% agreement and a weighted Cohen's kappa of 0.27 with ASE guidelines, while the clinician assessment had 32.7% agreement and a weighted Cohen's kappa of 0.06. Performance was consistent across patient subgroups stratified by sex, age, hypertension, diabetes, obesity, and coronary artery disease. Conclusion: Clinicians are often inconsistent in evaluating LVDD. We developed an AI pipeline that automates the clinical workflow of grading LVDD, which can contribute to improved diagnosis of heart failure.