Automated Deep Learning Pipeline for Characterizing Left Ventricular Diastolic Function

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 the potential for interobserver variability.

To characterize concordance in clinical evaluations of LVDD, we evaluated historical echocardiogram studies at 2 academic medical centers for variability between clinical diastology reported by clinicians and the 2016 and 2025 American Society of Echocardiography (ASE) guidelines. We developed a workflow of 8 deep learning (DL) models trained on over 155,000 studies to automate LVDD assessment. Model performance was evaluated on held-out test sets from 3 academic centers.

In 592 studies from Cedars-Sinai Medical Center (CSMC), DL demonstrated 80.4% agreement and a weighted Cohen kappa of 0.62 with the 2016 ASE guideline assessment. In contrast, the CSMC clinician evaluation had 52.2% agreement and a kappa of 0.46 with the 2016 guidelines. In the Stanford Healthcare (SHC) cohort of 1572 studies, DL had 66.7% agreement and a kappa of 0.33 with the 2016 guidelines vs clinical diastology, with 32.7% agreement and a kappa of 0.14. Our DL workflow showed strong concordance with the 2025 guidelines, with agreements of 81.8% in CSMC and 68.1% in SHC. In a third validation cohort of 204 studies from the University of Chicago, DL had 34.3% agreement with the 2025 guidelines.

Evaluating LVDD has high heterogeneity. We developed DL measurement-driven diastolic dysfunction, a pipeline of models that automates the clinical workflow of grading LVDD and can contribute to improved diagnosis of heart failure.