Automated Deep Learning Pipeline for Characterizing Left Ventricular Diastolic Function.

Area of Science:

• Cardiology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Left ventricular diastolic dysfunction (LVDD) assessment via echocardiography relies on complex algorithms and secondary characteristics, leading to interobserver variability.
• Current diagnostic methods for LVDD are laborious and lack a sole identifying metric, contributing to inconsistencies in clinical evaluations.

Purpose of the Study:

• To evaluate the concordance of clinical LVDD assessments with established guidelines.
• To develop and validate an artificial intelligence (AI) workflow for automating LVDD assessment using echocardiogram data.

Main Methods:

• Retrospective analysis of historical echocardiogram studies from two academic medical centers.
• Development of an 8-model AI workflow trained on over 155,000 echocardiogram studies for LVDD assessment.
• Performance evaluation of the AI workflow against 2016 American Society of Echocardiography (ASE) guidelines and clinician reports on distinct test sets.

Main Results:

• The AI workflow showed higher agreement (76.5% at Cedars-Sinai, 66.7% at Stanford) and Cohen's kappa (0.52 and 0.27, respectively) with ASE guidelines compared to clinician reports (48.5% and 32.7% agreement; 0.29 and 0.06 kappa).
• AI performance remained consistent across diverse patient subgroups, including variations in sex, age, and comorbidities like hypertension and diabetes.
• Clinician evaluations exhibited significant variability and lower concordance with ASE guidelines.

Conclusions:

• Clinician evaluations of LVDD demonstrate significant inconsistency.
• An AI pipeline was successfully developed to automate the grading of LVDD from echocardiograms.
• This automated AI approach has the potential to improve the accuracy and consistency of LVDD diagnosis, aiding in better heart failure management.