IDNoise: Resource-Aware Machine Learning-Based Noise and SNR Detection in Electrocardiogram Signals

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Wearable Electrocardiography (ECG) is susceptible to noise, particularly Motion Artifacts (MAs), which compromise signal integrity and interpretation.
• MAs present a significant challenge due to their complex time-frequency characteristics often overlapping with vital ECG signal components.
• Existing noise reduction methods struggle with the unpredictable nature of MAs, necessitating advanced detection and identification techniques.

Purpose of the Study:

• To develop and evaluate IDNoise, a Machine Learning (ML)-based system for the detection and identification of noise in ECG recordings.
• To leverage a comprehensive feature set for training ML models to differentiate noise types and estimate Signal-to-Noise Ratio (SNR).
• To assess the performance of IDNoise concerning accuracy, execution time, energy consumption, and memory usage for wearable applications.

Main Methods:

• Proposed IDNoise, an ML-based approach utilizing morphological, statistical, and concatenated features for noise analysis in ECG signals.
• Trained ML models for binary and 4-class noise type classification, and 7-class SNR level identification.
• Evaluated computational overhead including execution time, energy consumption, and memory usage, focusing on prediction and model loading phases.

Main Results:

• IDNoise achieved 80.52% accuracy and 80.44% F1-score in binary noise type classification using concatenated features.
• In 4-class noise classification, accuracy reached 67.91% with an F1-score of 67.89%; 7-class SNR identification yielded 44.80% accuracy and 44.57% F1-score.
• Feature extraction using concatenated features increased computational overhead (up to 7.5x), while prediction and model loading remained comparable to individual features.

Conclusions:

• IDNoise demonstrates a viable ML-based strategy for detecting and identifying noise in wearable ECG, improving signal reliability.
• The comprehensive feature set enhances classification performance but introduces computational trade-offs during feature extraction.
• IDNoise offers a promising solution for enhancing the robustness of wearable ECG monitoring systems against various noise interferences.