Transfer Learning Video Classification of Preserved, Mid-Range, and Reduced Left Ventricular Ejection Fraction in...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate assessment of left ventricular ejection fraction (EF) is crucial for patient management.
• Classifying EF into reduced (rEF), mid-range (mEF), and preserved (pEF) categories holds significant clinical importance.
• Traditional methods for EF assessment can be time-consuming and require specialized expertise.

Purpose of the Study:

• To evaluate the efficacy of end-to-end video classification using AutoML for categorizing left ventricular ejection fraction (EF).
• To assess the performance of AI models in distinguishing between reduced, mid-range, and preserved EF from echocardiographic recordings.
• To establish a proof of concept for automated EF classification using transfer learning.

Main Methods:

• Utilized Google Vertex AI AutoML for end-to-end video classification on echocardiographic datasets from the Standford EchoNet-Dynamic repository.
• Employed majority undersampling for dataset balancing and a 75/25 train-test split.
• Performed binary classification (rEF vs. not rEF; not pEF vs. pEF) and ternary classification (rEF, mEF, pEF).

Main Results:

• Binary classification of rEF vs. not rEF achieved high performance (ROC AUC 0.939, accuracy 0.863).
• Binary classification of not pEF vs. pEF demonstrated strong performance (ROC AUC 0.917, accuracy 0.829).
• Ternary classification showed lower performance, particularly for the mEF class. A PyTorch implementation confirmed feasibility.

Conclusions:

• End-to-end video classification using AutoML is a feasible approach for categorizing left ventricular ejection fraction (EF).
• The AI model demonstrated strong performance in binary classification tasks for identifying reduced or preserved EF.
• Further prospective clinical studies are warranted to evaluate this automated EF classification method.