Comparing subset selection methods in multi-lead electrocardiogram data

Area of Science:

• Biomedical Engineering
• Computational Cardiology
• Data Science

Background:

• Heart health complications are often detected via electrocardiogram (ECG) anomalies.
• Automated ECG summarization aids clinicians in timely and comprehensive patient assessment.
• Varying lead availability in different health settings necessitates flexible ECG analysis methods.

Purpose of the Study:

• To investigate subset selection algorithms for summarizing single-lead ECG, 12-lead ECG, vectorcardiogram (VCG), and VCG magnitude data.
• To compare the performance of seven CUR matrix decomposition algorithms, including oversampling techniques.
• To identify optimal ECG data representations and algorithms for improved diagnostic accuracy and computational efficiency.

Main Methods:

• Utilized the St. Petersburg INCART 12-lead Arrhythmia Database for analysis.
• Applied seven distinct CUR matrix decomposition algorithms for subset selection.
• Investigated both standard and oversampling approaches, including QR-based discrete empirical interpolation method (Q-DEIM) and extended DEIM (E-DEIM).

Main Results:

• The QR-based discrete empirical interpolation method (Q-DEIM) with 12-lead ECG data achieved high class detection among non-oversampling methods.
• The extended DEIM (E-DEIM) algorithm demonstrated superior overall performance using a lower-rank representation of VCG magnitude data.
• E-DEIM provided improved class detection with potential computational savings.

Conclusions:

• Subset selection is effective for summarizing diverse ECG representations.
• VCG magnitude data, when summarized using E-DEIM, offers a promising approach for efficient and accurate heart rhythm analysis.
• These summarized ECG representations can enhance subsequent diagnostic models and clinical decision-making for better patient outcomes.