Arrhythmia classification for non-experts using infinite impulse response (IIR)-filter-based machine learning and...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Cardiology

Background:

• Arrhythmias pose significant cardiovascular risks, necessitating accurate monitoring via electrocardiogram (ECG).
• Interpreting ECG data, especially 12-lead, is challenging due to signal noise and complexity, even for experts.
• Existing portable ECG monitors struggle with real-time interpretation and correlating rhythm data with patient conditions.

Purpose of the Study:

• To enhance ECG signal quality by removing noise and extracting features using a novel infinite impulse response (IIR) filter.
• To improve the understandability of ECG analysis for non-experts.
• To accurately classify cardiac rhythms using machine learning (ML) and deep learning (DL) models.

Main Methods:

• Utilized ECG data from the MIT-BIH database.
• Applied an infinite impulse response (IIR) filter for noise reduction and feature extraction.
• Employed hyperparameter (HP)-tuning for ML classifiers and fine-tuning (FT) for DL models, including DenseNet-121.
• Evaluated different filters for arrhythmia categorization and assessed accuracy changes.

Main Results:

• The proposed IIR filter effectively improved ECG signal quality.
• Machine learning and deep learning models demonstrated high performance in rhythm classification.
• DenseNet-121 achieved 99% accuracy without fine-tuning and 99.97% accuracy with fine-tuning.

Conclusions:

• The IIR filter significantly enhances ECG signal processing for improved arrhythmia detection.
• Fine-tuned deep learning models, particularly DenseNet-121, offer highly accurate cardiac rhythm classification.
• This research contributes to more reliable and accessible ECG analysis for better patient outcomes.