Assessment and comparison of different methods for heartbeat classification

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Electrocardiogram (ECG) signal analysis, particularly QRS complex assessment, is crucial for diagnosing cardiac dysfunctions.
• Existing automatic heartbeats classification methods lack direct comparability due to varied features and datasets.
• A comparative analysis is needed to evaluate different classification techniques' effectiveness on ECG data.

Purpose of the Study:

• To comparatively study the learning capacity and classification abilities of four distinct methods: Kth nearest neighbour rule, neural networks, discriminant analysis, and fuzzy logic.
• To assess these methods using 26 morphological ECG parameters, including amplitude, area, interval durations, and QRS vector.
• To evaluate performance across five ventricular complex types: normal, premature ventricular contractions, left bundle branch block, right bundle branch block, and paced beats.

Main Methods:

• Four classification algorithms (Kth nearest neighbour, neural networks, discriminant analysis, fuzzy logic) were applied.
• Analysis utilized 26 morphological parameters derived from ECG signals.
• Performance was evaluated using three learning set strategies: global, basic, and local (consecutive and random).

Main Results:

• Classification methods achieved high accuracies when trained on local learning sets specific to each patient.
• A small basic learning set, though balanced, demonstrated reduced classification ability.
• The global learning set, encompassing data from all patients, yielded the poorest classification results.

Conclusions:

• Local learning sets are most effective for accurate patient-specific ECG classification.
• The choice of learning set significantly impacts the performance of cardiac rhythm classification algorithms.
• Further research should focus on optimizing learning set strategies for robust ECG analysis.