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

Noise is a significant challenge in wearable Electrocardiography (ECG), as it can distort the ECG waveform and lead to inaccurate signal interpretation. Among noise sources, Motion Artifcats (MA)s are particularly difficult to mitigate due to their unpredictable time-frequency characteristics. Unlike electromagnetic interference, MAs often overlap with critical spectral components of the ECG signal, making their reduction in preprocessing especially challenging. In this work, we propose IDNoise, a Machine Learning (ML)-based approach for detection and identification of noise in ECG recordings. Our method leverages a comprehensive feature set, including morphological, statistical, and concatenated features, to train ML models capable of distinguishing various noise types and estimating their Signal to Noise Ratio (SNR). We evaluate the proposed algorithms in terms of execution time, energy consumption, and memory usage, which are critical resource usage metrics when designing solutions for wearable devices. We demonstrate that IDNoise, when using concatenated features, can detect the noise type in binary classification with an accuracy of 80.52% and an F1-score of 80.44%. It can also detect the noise type in 4-class classification with an accuracy of 67.91% and an F1-score of 67.89%, and identify the SNR level in 7-class classification with an accuracy of 44.80% and an F1-score of 44.57%. While the use of concatenated features necessitates, during the feature extraction phase, an increase in computational overhead of a factor up to 7.5 times (execution time, energy consumption, and memory usage). The computational overhead for prediction and loading the models remain comparable to that of individual features.