Keywords
Noise invariance; Noise tolerance; Wearable sensors; Electrocardiogram; Automatic diagnosis
Abstract
Sensors have been widely used in various data acquisition systems, especially for medical applications. However, once developed for wearable use, these have suffered from various types of noise which greatly degrade data quality and consequently, the reliability. Low data quality is a major obstacle for computer-based diagnosis. Thus, the noise tolerance ability plays a crucial role in wearable sensor based data acquisition and analysis. This work proposes a novel method: noise-invariant component analysis (N-ICA), to expose the influence of noise on this data and provides noise removal and dimensionality reduction. The proposed N-ICA based approach extracts information from data which undergoes minimal change with noise and directly shows the extent to which the true information has been corrupted by noise. This work also implements a simulated wearable sensor based ECG automatic diagnosis system together with a noise generator to validate N-ICA noise tolerant enhancement. Test data is selected from the MIT-BIH Arrhythmia Database. The simulated ECG monitoring system achieves 99.42% accuracy in classifying eight types of heartbeats. Experimental results demonstrate that the signal-to-noise ratio is improved by applying N-ICA based on ECG data contaminated by five noise sources. QRS detection accuracy is also improved to above 95% under the highest noise level tested. Dimensionality reduction reduces the data to 6.5% of the original size. Finally, diagnostic accuracy of four different classifiers is significantly improved when applied in our simulated ECG automatic monitoring system.
Citation
Chen K, Powers LS and Roveda JM. Noise-Invariant Component Analysis for Wearable Sensor based Electrocardiogram Monitoring System. SM J Biomed Eng. 2018; 4(1): 1025.