Multilevel Signal Quality Assessment and Noise Type Detection Framework for Quality-Aware Wearable ECG Monitoring Device

Abstract

Wearable devices offer continuous monitoring of electrocardiogram (ECG) signals for arrhythmia detection with limited computing resources. Automatic signal quality assessment (SQA) is crucial for enhancing the reliability of such devices and minimizing false alarms. In addition, identifying noise types can improve the effectiveness of noise-specific denoising strategies in such limited resource-based devices. This study proposes a lightweight deep learning framework for multilevel ECG SQA and noise-type classification suitable for resource-constrained wearable monitoring devices. The model architecture with a suitable activation function is optimized through extensive comparative performance analysis. The proposed approach effectively classifies clean ECG (CL-ECG), ECG with low-frequency noise (LFN-ECG), ECG with high-frequency noise (HFN-ECG), ECG with mixed-frequency noise (MFN-ECG), and unacceptable ECG (UN-ECG). The test ECG segments are generated from the standard MIT-BIH arrhythmia (MIT-BIHA) database, PhysioNet Challenge 2011 (PCCC-2011) database, and PhysioNet Challenge 2021 (PCCC-2021) dataset. The optimized model demonstrates an average F1-score of 98.05% for the MIT-BIHA. For completely unknown test datasets, the F1-score of 95.93% and 94.20% are reported for PCCC-2011 and PCCC-2021, respectively. The proposed framework is implemented on a Raspberry Pi platform to assess its feasibility in a resource-limited platform. The optimized trained model with a size of only 435 kB effectively classifies real-time ECG signal quality with an average computation time of 510.19 ms on a Raspberry Pi. The detailed analysis of the results confirms that the proposed approach classifies ECG signal quality and noise types in real-time on a limited computing resource-based platform which makes it well-suited for wearable ECG monitoring devices.