Comparison of multigroup logistic and linear discriminant ECG and VCG classification

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Electrocardiogram (ECG) and vectorcardiogram (VCG) are crucial for diagnosing cardiac conditions.
• Accurate classification of cardiac abnormalities like hypertrophy and myocardial infarction (MI) is essential for patient management.
• Comparing statistical models for ECG/VCG analysis can optimize diagnostic utility.

Purpose of the Study:

• To evaluate and compare the performance of logistic (LOG) and linear discriminant analysis (LDA) for ECG and VCG data.
• To assess the accuracy and reliability of these models in classifying various cardiac conditions.
• To determine if LOG or LDA models offer superior diagnostic capabilities for routine clinical practice.

Main Methods:

• Utilized a large validated database with learning (N=2446) and test (N=820) sets.
• Applied LOG and LDA models to both 12-lead ECG and 3-lead VCG data.
• Classified seven groups: normal, left, right, biventricular hypertrophy, and anterior, inferior, combined MI.

Main Results:

• LOG models achieved up to 3% higher total accuracy than LDA, with similar rates between 12-lead and 3-lead analyses (67-70% for 7 groups, 77-84% for 5 groups).
• LDA showed higher specificity for normal subjects but lower sensitivity for specific conditions compared to LOG.
• LOG models demonstrated better fit to patient data and less overconfidence than LDA, particularly for anterior MI classification.

Conclusions:

• LOG models provide a marginal but significant improvement in classification accuracy (1-3%) over LDA for ECG and VCG data.
• The enhanced reliability of LOG models is a key advantage, potentially improving diagnostic utility in clinical settings.
• These findings suggest LOG analysis can be a valuable tool for routine ECG and VCG interpretation.