Improved sleep–wake and behavior discrimination using MEMS accelerometers

https://doi.org/10.1016/j.jneumeth.2007.03.007Get rights and content

Abstract

State of vigilance is determined by behavioral observations and electrophysiological activity. Here, we improve automatic state of vigilance discrimination by combining head acceleration with EEG measures. We incorporated biaxial dc-sensitive microelectromechanical system (MEMS) accelerometers into head-mounted preamplifiers in rodents. Epochs (15 s) of behavioral video and EEG data formed training sets for the following states: Slow Wave Sleep, Rapid Eye Movement Sleep, Quiet Wakefulness, Feeding or Grooming, and Exploration. Multivariate linear discriminant analysis of EEG features with and without accelerometer features was used to classify behavioral state. A broad selection of EEG feature sets based on recent literature on state discrimination in rodents was tested. In all cases, inclusion of head acceleration significantly improved the discriminative capability. Our approach offers a novel methodology for determining the behavioral context of EEG in real time, and has potential application in automatic sleep–wake staging and in neural prosthetic applications for movement disorders and epileptic seizures.

Introduction

There is a long history of determining the state of vigilance for humans (Broughton, 1999) or animals (Robert et al., 1999) using EEG criteria. Nevertheless, there is considerable uncertainty when determining the state of vigilance using EEG alone. Therefore, EEG is typically augmented with simultaneous visual behavioral monitoring, and/or the incorporation of electromyogram (EMG) or electrooculogram (EOG) recordings.

We are developing the technical capability for discriminating state of vigilance in real time. Our objective is to provide contextual input for seizure prediction and control. It is readily accepted that the dynamics of EEG change remarkably in the different stages of wakefulness and sleep (Niedermeyer, 1999). Yet much of the current work in seizure detection and prediction (Mormann et al., 2005) focuses on statistical or dynamical changes of the EEG with respect to a baseline defined without regard to state. Our end goal is to implement state-dependent seizure detection and control in medical devices suitable for human implantation during the activities of daily living, where continuous video monitoring would not be feasible and EMG or EOG electrodes might be invasive or cumbersome.

In this study, we explored combining head acceleration measurements with EEG in order to improve our ability to discriminate state of vigilance in rodent experiments. We incorporated biaxial dc-sensitive microelectromechanical system (MEMS) accelerometers into the head-mounted preamplifier circuit used for EEG recording. We used combined EEG and behavioral video to establish training and validation data sets, and then used EEG features with and without accelerometer features in a multivariate linear classifier. We tested a broad range of EEG feature sets based on those used in the recent literature for state discrimination in rodents (Robert et al., 1999). Our approach offers a novel methodology for determining the behavioral context of EEG in real time. Preliminary results of this investigation were previously reported (Peixoto et al., 2004, Sunderam et al., 2005).

Section snippets

Surgical implantation and data acquisition

Video (3 fps including visual and infrared sensitivity), EEG and head acceleration were recorded continuously from adult male Sprague-Dawley rats 200–300 g in weight (Harlan Ltd., NY) using a custom-made electronics and acquisition system (Labview, National Instruments Ltd.).

Head-mounted preamplifiers were constructed with integrated dc-sensitive biaxial MEMS accelerometers (ADXL 311, Analog Devices Ltd.; sensitivity of 312 mV per g, where g = acceleration due to gravity, at the reference voltage of

Manual scoring statistics

For each animal, the fraction of epochs in the training set manually scored as each individual state is recorded in Table 1. Not counting Indeterminate epochs, the rats spent about 60% of the time in sleep, and over 80% of sleep in SWS. These values compare well with the literature (Gervasoni et al., 2004, Antle and Mistlberger, 2005) which indicates that this data set is representative of typical circadian behavior for rats.

Motion signatures of states

The animal's overall activity level was reflected in the total power

Discussion

We have demonstrated that the addition of head acceleration measurements to EEG spectral measurements significantly improves LDA use to automatically classify sleep–wake and behavioral state. With EEG alone, we observed wide variability in discrimination performance based on the set of EEG variables used (Costa-Miserachs et al., 2003, Gervasoni et al., 2004, Louis et al., 2004, Sunderam et al., 2005). The addition of acceleration improved and equalized their performance, and rendered the choice

Conclusion

We have shown significant improvement in state discrimination accuracy afforded by the addition of acceleration measurements to EEG. MEMS accelerometers are lightweight (0.15 g) and inexpensive, less invasive than EMG or EOG, and the modifications in circuitry over that required for EEG acquisition are minimal. We intend to use this improved behavioral state detection capability to formulate and inform detection and feedback control algorithms for treating seizures.

Acknowledgements

This work was funded by National Institutes of Health grants R01EB001507, K02MH01493 and R01MH50006.

References (27)

  • P.A. Tiba et al.

    Effects of early handling on basal and stress-induced sleep parameters in rats

    Brain Res

    (2003)
  • N.C. Aggelopoulos et al.

    Responses of neurones of the rat suprachiasmatic nucleus to retinal illumination under photopic and scotopic conditions

    J Physiol

    (2000)
  • M. Antle et al.

    Circadian rhythms

  • Cited by (35)

    • Noninvasive three-state sleep-wake staging in mice using electric field sensors

      2020, Journal of Neuroscience Methods
      Citation Excerpt :

      Often this discrepancy was a single 10-second epoch, but occasionally more. REM-related scoring error is also common in EEG/EMG sleep-wake scoring methods partially due to the low occurrence (typically 5–15% of total sleep time) (Younes et al., 2016; Younes and Hanly, 2016) and placement of EEG electrodes (Sunderam et al., 2007). However, accuracy in identifying REM sleep was not affected by the quality (i.e. signal to noise ratio) of the data.

    • Opportunities for improving animal welfare in rodent models of epilepsy and seizures

      2016, Journal of Neuroscience Methods
      Citation Excerpt :

      Several approaches are used, either alone or in combination. Video recording or other measurements such as micro-electro-mechanical systems (MEMS), can detect movements associated with seizures (Sunderam et al., 2007). Automated observational systems (e.g. Van de Weerd et al., 2001) can be used to analyse the behavioural profile of animals following seizure induction (Riljak et al., 2014) and methods are currently being developed to allow automated recordings in the home cage environment4.

    • EEG gamma frequency and sleep-wake scoring in mice: Comparing two types of supervised classifiers

      2010, Brain Research
      Citation Excerpt :

      In contrast to our empirical approach, the attempt to use PCA for systematic parameter reduction and selection of the strongest stage indicators was not successful in the context of supervised scoring algorithms (but see: Jobert et al. (1994) and Martinez and Kak (2001)). LDA (Molinari et al., 1984; Sunderam et al., 2007) and Tree (Hanaoka et al., 2001) and combinations of both (Anderer et al., 2005) have been used previously for sleep–wake scoring in humans. To our knowledge, however, they have not been adapted for sleep scoring in mice or rats.

    View all citing articles on Scopus
    View full text