Prediction of left ventricular systolic dysfunction in left bundle branch block using a fine-tuned ECG foundation...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Left bundle branch block (LBBB) is a significant ECG finding associated with left ventricular systolic dysfunction (LVSD).
• Early detection of LVSD is critical for patient outcomes, but echocardiography is not always readily accessible.
• Standard diagnostic methods for LVSD can be time-consuming and may not be immediately available in all clinical settings.

Purpose of the Study:

• To develop and evaluate a fine-tuned ECG foundation model (FM) for enhanced LVSD detection in LBBB patients.
• To compare the performance of the proposed ECG-FM against conventional deep learning models.
• To identify key ECG features predictive of LVSD in LBBB patients using explainability techniques.

Main Methods:

• Retrospective multicenter analysis of 2,031 paired ECG-echocardiographic datasets from 892 LBBB patients.
• Fine-tuning a pre-existing ECG foundation model for LVSD prediction.
• Comparison with baseline deep learning models: Fully Convolutional Network (FCN), LSTM-FCN, ResNet, and InceptionTime.
• Utilized DeepLIFT analysis for feature importance assessment.

Main Results:

• The fine-tuned ECG-FM achieved an accuracy of 0.758, sensitivity of 0.771, and AUROC of 0.807, outperforming baseline models.
• Sequential partial fine-tuning yielded the highest sensitivity (0.787), enhancing screening capabilities.
• DeepLIFT analysis highlighted QRS complex and T wave features in leads V1-V4 as crucial predictive factors.

Conclusions:

• The fine-tuned ECG-FM significantly improves LVSD detection in LBBB patients.
• This AI-driven approach offers potential for earlier clinical diagnosis when echocardiography is unavailable.
• The model's ability to identify critical ECG features may enhance clinical management and improve patient outcomes.