oa Energy-efficient data reduction techniques for EEG wireless body sensor networks

With recent advances in signal processing and very-low-power wireless communications, wireless body sensor networks (WBSNs) are gaining wide popularity. A WBSN consists of multiple miniaturized sensors that are placed on the person's body and are capable of measuring and communicating different physiological signals over time. This study focuses on WBSNs that rely on electroencephalogram (EEG) signals. EEG signals measure the electrical brain activity through a collection of non-invasive wireless sensors placed on a patient's scalp. Two applications are studied: the development of brain computer interfaces (BCIs), and the detection of epileptic seizures. A BCI is a direct interface between the brain and a machine. It can be used for purposes such as helping a patient perform a task by thought only, i.e. without performing any motor actions. In such a case, the BCI has to detect the presence of specific command signals in the EEG signals. A WBSN has the advantage of being minimally obtrusive to the patient. This is because the signals are transmitted wirelessly from the person's body; a person can therefore move freely without worrying about surrounding wires. However, in WBSN applications, the energy available in the battery-powered sensors is limited. Different solutions to minimize the number of computations carried out and the amount of data transmitted by the sensor are therefore highly desired. In this study, we present computationally-efficient data reduction techniques to reduce the energy consumption at the sensor node while keeping the salient information in the EEG signals. To efficiently compress EEG signals at the sensor node, we propose the use of a compressed sensing (CS) framework. The proposed CS scheme is simple, nonadaptive and yields higher energy efficiency than existing frameworks. To obtain a high compression ratio, our CS framework exploits not only the temporal correlation within EEG signals in each channel as is the case in existing frameworks, but also the inter-correlation amongst different EEG channels. When applied to a simple BCI system, our proposed framework resulted in important energy savings (up to 60%) at the expense of a slightly reduced classification accuracy. Existing BCIs require all the EEG signals as input. Therefore, the EEG signals must be reconstructed as perfectly as possible at the receiver side. For seizure detection however, the main aim is not to reconstruct the EEG signals but to detect the occurrence of a seizure. In addition to the above CS technique, we examined different data reduction techniques at the sensor side of an EEG seizure detection system. The extraction and transmission of certain features of the EEG signals were found to yield best results. The performance of these techniques was evaluated based on power consumption and seizure detection efficacy. Experimental results showed that by performing low-complexity feature extraction and transmitting only the features that are pertinent to seizure detection, considerable overall energy is saved. The battery life of the system is increased 14 times relative to the conventional approach of transmitting all the original EEG signals, while the same seizure detection performance is maintained (94.1% sensitivity).