Abnormality classification from electrocardiograms with various lead combinations | JoVE Visualize

Area of Science:

Cardiology
Biomedical Engineering
Artificial Intelligence in Medicine

Background:

Cardiovascular diseases are a leading cause of death, necessitating cost-effective and accurate early diagnosis of cardiac abnormalities.
Developing generalized models for identifying multiple cardiac abnormalities from diverse electrocardiogram (ECG) datasets, including 12-lead and reduced-lead signals, presents significant challenges due to data divergence and label corruption.

Purpose of the Study:

To build robust and accurate models capable of classifying 30 types of cardiac abnormalities from various combinations of ECG leads.
To address data divergence, label corruption, and distribution uncertainty in ECG datasets for improved model generalization.

Main Methods:

A preprocessing workflow was implemented to mitigate data divergence across different ECG datasets.
A squeeze-and-excitation deep residual network was employed as the base model to capture lead-wise relationships in ECG signals.
Novel strategies including cross-relabeling, sign-augmented loss, pos-if-any-pos ensemble, and dataset-wise cross-evaluation were utilized to handle corrupted labels and data distribution uncertainties.

Main Results:

The proposed framework achieved challenge metric scores of 0.57, 0.59, 0.59, 0.58, and 0.57 for 12-, 6-, 4-, 3-, and 2-lead ECG signals, respectively.
An averaged challenge metric score of 0.58 was obtained across all lead versions in the Physionet/Computing in Cardiology Challenge 2021.
The models demonstrated accurate identification of 30 ECG abnormalities on hidden test data, validating the framework's effectiveness.

Conclusions:

The developed models, trained on large datasets, accurately classify 30 ECG abnormalities using various lead combinations.
The proposed framework effectively addresses challenges in ECG data, leading to robust models for cardiac abnormality detection.
The performance on hidden test data confirms the significant potential of the proposed approaches for real-world clinical applications.