A neural algorithm for the non-uniform and adaptive sampling of biomedical data

• Published in: Computers in biology and medicine Vol. 71; pp. 223 - 230
• Main Author: Mesin, Luca
• Format: Journal Article
• Language: English
• Published: United States Elsevier Ltd 01.04.2016; Elsevier Limited
• Subjects: Accelerometer; Accelerometry - methods; Algorithms; Automatic Data Processing - methods; Compressive sensing; Conflicts of interest; Data compression; Decision making; Electrocardiography; Electrocardiography (ECG); Electrocardiography - methods; Electroencephalography; Electroencephalography (EEG); Electroencephalography - methods; Electromyography (EMG); Internal Medicine; Models, Theoretical; Non-uniform sampling; Nyquist limit; Other; Physiology; Real time; Self evaluation; Sensors; Variables; Compressive sensing; Electroencephalography (EEG); Electromyography (EMG); Electrocardiography (ECG); Nyquist limit; Accelerometer; Non-uniform sampling
• ISSN: 0010-4825; 1879-0534
• DOI: 10.1016/j.compbiomed.2016.02.004

Body sensors are finding increasing applications in the self-monitoring for health-care and in the remote surveillance of sensitive people. The physiological data to be sampled can be non-stationary, with bursts of high amplitude and frequency content providing most information. Such data could be sampled efficiently with a non-uniform schedule that increases the sampling rate only during activity bursts. A real time and adaptive algorithm is proposed to select the sampling rate, in order to reduce the number of measured samples, but still recording the main information. The algorithm is based on a neural network which predicts the subsequent samples and their uncertainties, requiring a measurement only when the risk of the prediction is larger than a selectable threshold. Four examples of application to biomedical data are discussed: electromyogram, electrocardiogram, electroencephalogram, and body acceleration. Sampling rates are reduced under the Nyquist limit, still preserving an accurate representation of the data and of their power spectral densities (PSD). For example, sampling at 60% of the Nyquist frequency, the percentage average rectified errors in estimating the signals are on the order of 10% and the PSD is fairly represented, until the highest frequencies. The method outperforms both uniform sampling and compressive sensing applied to the same data. The discussed method allows to go beyond Nyquist limit, still preserving the information content of non-stationary biomedical signals. It could find applications in body sensor networks to lower the number of wireless communications (saving sensor power) and to reduce the occupation of memory. The paper proposes an innovative real time method to under-sample non-stationary data, providing applications to four different biomedical signals: electromyogram (EMG), electrocardiogram (ECG), electroencephalogram (EEG) and body acceleration. An example of application is shown in Figure 1, where an EMG is sampled under the Nyquist limit, adapting the sampling schedule to the data. The sampling rate is automatically increased during bursts of activity of the muscle and lowered when when the EMG has small amplitude (reflecting cross-talk from nearby muscles or noise). The Power Spectral Density (PSD) is fairly well represented by the adaptive under-sampling till the highest frequency components, even under a drastic reduction of the number of samples. The method outperforms both a uniform under-sampling and compressive sensing (CS) applied to the same data with the same compression ratio. The algorithm opens interesting perspectives for potential applications in body sensor networks (which are finding increasing applications, e.g. in self-monitoring and for the surveillance of sensitive people). Specifically, it could allow to lower the number of wireless communications (saving sensor power) and to reduce the occupation of memory. [Display omitted] •Algorithm to under-sample data with a frequency lower than Nyquist limit.•An adaptive neural predictor selects the subsequent samples to be measured.•Example applications: EMG, ECG, EEG and acceleration data.•The method outperforms uniform sampling and compressive sensing.•Many potential applications to save energy and to reduce the memory storage.