[Paper Review] ECG beats classification using waveform similarity and RR interval

Ahmad Khoureich Ka|arXiv (Cornell University)|Jan 10, 2011

ECG Monitoring and Analysis9 references21 citations

TL;DR

This paper proposes a patient-adaptive ECG beat classification method using waveform similarity and RR interval features, employing discrete wavelet transform for denoising and beat reduction. The approach achieves 97.52% classification accuracy on 46 MIT-BIH arrhythmia database records across six beat types, with improved performance (98.53%) when excluding records with high RR interval variability.

ABSTRACT

This paper present an electrocardiogram (ECG) beat classification method based on waveform similarity and RR interval. The purpose of the method is to classify six types of heart beats (normal beat, atrial premature beat, paced beat, premature ventricular beat, left bundle branch block beat and right bundle branch block beat). The electrocardiogram signal is first denoised using wavelet transform based techniques. Heart beats of 128 samples data centered on the R peak are extracted from the ECG signal and thence reduced to 16 samples data to constitute a feature. RR intervals surrounding the beat are also exploited as feature. A database of annotated beats is built for the classifier for waveform comparison to unknown beats. Tested on 46 records in the MIT/BIH arrhythmia database, the method shows classification rate of 97.52%.

Motivation & Objective

• To develop a low-complexity, patient-adaptive ECG beat classification system that improves accuracy by leveraging individual patient morphology.
• To address the challenge of morphological variability in ECG beats across and within patients, which hinders traditional automated classification.
• To reduce noise in ECG signals using discrete wavelet transform for improved beat feature extraction.
• To build a dynamic beats database from manually annotated beats for real-time classifier adaptation.
• To achieve high classification accuracy using a similarity-based classifier without relying on neural networks or complex models.

Proposed method

• ECG signals are denoised using a combination of baseline wander cancellation and soft thresholding via discrete wavelet transform (DWT).
• Heartbeats are extracted as 128-sample segments centered on the R peak and reduced to 16-sample features using DWT for compact representation.
• RR intervals before and after each beat are used as additional features to enhance classification.
• A patient-specific beats database is constructed from the first 5 minutes of manually annotated beats in each ECG record.
• Beats are classified by clustering based on waveform similarity (threshold ≥ 0.95) and assigning labels based on the closest match in the beats database.
• The similarity function compares normalized 16-sample waveforms, and classification relies on patient-specific knowledge accumulation.

Experimental results

Research questions

• RQ1 Can a patient-adaptive ECG beat classifier achieve high accuracy using only waveform similarity and RR interval features?
• RQ2 How does the integration of patient-specific annotated beats improve classification performance over generic models?
• RQ3 To what extent does RR interval variability affect misclassification rates in ECG beat classification?
• RQ4 Can wavelet-based preprocessing effectively reduce noise while preserving diagnostic morphology for beat classification?
• RQ5 How does the proposed method compare in accuracy and computational cost to neural network or fuzzy logic-based approaches?

Key findings

• The proposed method achieved an overall classification accuracy of 97.52% across 46 records from the MIT-BIH arrhythmia database.
• Excluding records with high RR interval variability (203, 222, 223), the accuracy increased to 98.53%.
• Record 203 had the lowest accuracy (76.00%), primarily due to morphological confusion between normal and atrial premature beats caused by RR interval fluctuations.
• The method outperformed several existing approaches in accuracy for six beat types, including 99.51% for ICA and 98.35% for FHNN, though with different test conditions.
• The classifier demonstrated low computational cost and scalability through incremental learning via patient-specific beat database updates.
• Misclassification was often linked to R-peak misalignment in the MIT-BIH annotated files, suggesting a need for improved R-peak detection in future work.