Extracting signals from noisy single-channel EEG stream for ubiquitous healthcare applications

A large number of ubiquitous healthcare systems and applications have been recently introduced as a result of breakthroughs in technologies such as wireless sensors and BAN (Body Area Network). However, most of the previous related work has focused on the technology platform and service architecture for u-healthcare, while sensor data acquisition and handling have not received much attention. The sensors' readings are typically unreliable and contain a lot of noise, especially when they are used in uncontrolled environments such as ubiquitous healthcare systems. Quality of service cannot be guaranteed without the proper handling of noise. The problem is exacerbated for EEG signals; because the signal-to-noise ratio is especially low, the number of channels is limited, and noise (mostly ocular artifacts) removal should be done online. In this paper, we introduce a method, called Online SSA, in order to address this problem. Online SSA extends the conventional offline SSA by incorporating the rank-1 modification technique to incrementally update the singular spectrum of the noise model. We validated the proposed method using real EEG data generated from a single channel EEG device. We showed that our algorithm does not require pre-training; it rapidly builds up an accurate noise model from initial user feed. Moreover, we compared the effects of varying data acquisition parameters. Finally, we compared the power spectrum densities of the result EEG signals to the clean baseline EEG to demonstrate the effectiveness of the proposed method.