Electrocardiogram analysis using a combination of statistical, geometric, and nonlinear heart rate variability...

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Electrocardiogram (ECG) classification using heart rate variability (HRV) analysis is crucial for diagnosing cardiac conditions.
• Current limitations exist in fully understanding HRV analysis for diagnosing diverse cardiac issues.
• Combining linear and nonlinear HRV features is proposed to enhance classification accuracy.

Purpose of the Study:

• To evaluate a proposed combination of linear and nonlinear HRV features for improved ECG classification.
• To compare the performance of various machine learning algorithms in HRV analysis.
• To determine the optimal analysis period for nonlinear HRV features.

Main Methods:

• Extracted 11 HRV features (e.g., SDNN, RMSSD, HTI, ApEn) from 5-minute R-R interval recordings of 100 patients across four conditions.
• Utilized seven machine learning algorithms (K-means, EM, C4.5, Bayesian Network, ANN, SVM, RF) for binary and four-class classification.
• Evaluated feature relevance using 1-Rule and C4.5, and assessed the impact of analysis period T on nonlinear features.

Main Results:

• Random Forest (RF) achieved the highest accuracy: 99.7% for two-class and 99.6% for four-class classification.
• Key features identified include HTI, pNN20, RMSSD, ApEn3, ApEn4, and SFI, with HTI being particularly effective.
• Using five periods for nonlinear feature extraction improved accuracy from 70% to 99%.

Conclusions:

• A combination of 11 linear and nonlinear HRV features, with nonlinear features extracted over five periods, yields high classification accuracy.
• The RF algorithm is highly effective for both binary and multiclass HRV classification.
• Further research should focus on identifying optimal feature sets for specific cardiac disorders.