Up to the year 2000, BCI was almost exclusively pursued utilising electrical measures of cortical activity - mostly from the intact scalp, i.e. using EEG (Wolpaw, et al., 2000). Recently, more BCI modalities have arisen such as MEG-based BCIs and the metabolic oriented fMRI-based BCIs, although EEG-based systems are still the dominant choice by research groups (Mason, et al., 2007). This is perhaps due to the practicality, portability, bedside usability of EEG, and due to the technological development it has received since Berger's pioneering work (Berger, 1929). Subsequently Coyle introduced NIRS-based BCI to the BCI community (Coyle, et al., 2004) developing a single\dual-channel custom-made instrument for BCI. One justification for an fMRI-based BCI and indeed MEG-based BCI were that they are postulated as being potential training grounds for BCIs. In particular MEG-based BCIs were directly comparable to EEG-based BCIs since the recorded features of cerebral effort were similar and thus interchangeable (Mellinger, et al., 2007; Waldert, et al., 2008). Ultimately a BCI should be portable and have use at the bedside, which makes MEG and fMRI based systems currently impractical, but are nonetheless useful additions to BCI research. Moreover, fMRI has a correlate with NIRS-based systems (in HbR) and potentially if needed, could also be an interchangeable training ground (or perhaps simply for localisation) for portable NIRS-based BCI. For this project, fMRI work on functional motor activity was conducted to reveal neural substrates that were both independently and simultaneously activated (see Chapter 6). This millimeter spatial resolution offered by fMRI could be used to determine an intended user's cortical regions of maximum activity and this information could then be utilised in a NIRS-based BCI. Nevertheless for single-trial NIRS, specific events may vary to those determined from grand averages established from fMRI, although fMRI can and has been used in a single-trial implementation (Yoo, et al., 2004; Weiskopf, et al., 2007).
4.3.1 Why NIRS?
A NIRS-based BCI offers many potential advantages for BCI and these will be mentioned shortly. However, the most commonly reported limitation of a NIRS-based BCI is that it is intrinsically limited by the time taken for a functional response to be elicited (~5 seconds). Thus, typically a functional response is determined by a large increase in HbO2 and a significant but smaller decrease in HbR, taking approximately 4-8 seconds to peak, with a similar time to return to baseline. Thus, the potential bit rate for a single channel NIRS-BCI using this haemodynamic feature is ~6 bits/min. However, there are four key points (but also other advantages listed later in this section) in favour of this relatively slower technique:
1) Sufficient Speed: Although in general greater speed is favoured in terms
of computer processing and essentially the time it takes to carry out a computational task, these desires may not always be similar for such an essential communication tool as a BCI. Recently, Allison et al reported how a 0.96bit/min device can still allow affective communication and they review that even slower systems have been continued to be used by patients (Allison, et al., 2008);
2) Alternative Channels - choice: Continuing from the previous point of
speed is the element of choice. Although a certain BCI may emerge with the highest bit rate (e.g. the SSVEP-BCIs with up to 68 bits/min (Gao, et al., 2003)), some require the application of external stimuli (extraneous) which can be a cause of irritation for the user. Thus, choice should feature in the selection of a BCI that perhaps reflects a user's preference. A NIRS- based BCI offers an alternative passive (endogenous) channel for BCI research;
3) Specificity: Certain neural substrates could become impaired as a result of
the particular disability of the user such as severe cerebral palsy (Daly, et al., 2008). As such NIRS offers high parameter specificity of the haemoglobin species being measured (Obrig, et al., 2003) and can exhibit distinct (independent) functional activations such as motor and visual evoked responses (Plichta, et al., 2006). Thus, as reported by Daly et al, BCI techniques most effective will depend on the underlying CNS abnormality. They also specify that certain signals (e.g. SMRs) may be absent in some users due to the cerebral impairment and so alternative signals would be required. Thus, NIRS-based BCIs may be able to provide this alternative. This again links to the idea of variety and choice for BCI. Furthermore, a hybrid EEG-NIRS device could be effective in this case to establish alternatives, since NIRS optical signals would not interfere with the EEG recording, i.e. it is electrically isolated using photonic energy for interrogation of cerebral tissue;
4) Outlook: Finally, it should be noted that there are optical signals found
from invasive studies (Thompson, et al., 2003) that are localised to cortical columns and others reporting high temporal responses on the order of milliseconds (Obrig, et al., 2003). These have been mooted as being measurable non-invasively using a 1000 averages or more (Gratton, et al., 1997; Wolf, et al., 2004) for a 0.07% change in signal intensity (Franceschini, et al., 2004) although other groups have reported an
inability to reproduce some of these results (Obrig, et al., 2003). However, it is restated here that further advances in technological development to enhance the SNR of in vivo NIRS (ivNIRS) to detect these signals related to neuronal cell swelling, may provide an avenue for a faster optical BCI. 0)
As an aside, it should be noted that many researchers state that NIRS has a low temporal resolution - it does not. Temporal resolution for NIRS is on the order of ~100msec, but what is meant\intended really is that NIRS has a haemodynamic feature that has a large time constant - i.e. a concentration change in HbO2 and in HbR takes 4-8 seconds to peak. In addition, there are potentially earlier transient features that could be utilised such as the elusive initial dip, sometimes found in fMRI (Buxton, 2001; Buxton, et al., 2004) but more research is needed to investigate this in single-trial NIRS.
Optical BCIs have further advantages such as: 1) practicality, 2) providing a low-cost alternative relative to other metabolic BCIs (e.g. fMRI), 3) portability (including potential wireless implementations Muehlemann, et al., 2008), 4) suitability for the bedside (required for BCI), 5) being non-invasive, 6) suitability for long-term use, 7) being safe, 8) having a high temporal resolution (~100msec), 9) user-friendliness (sensors don't require gel as in EEG), and 10) requiring less training (Coyle, et al., 2007) since the functional activity is from a more natural (direct) mental procedure compared to the intense learning required in some BCIs - e.g. SCP-based BCIs. Months of training can be required for such devices, leading to frustration and perhaps abandonment of the device since the user doesn't normally have control over these signals - they are learned (Spinney, 2003).
4.3.2 Justification for multichannel-OBCI
Many commercial multichannel NIRS instruments are available (Wolf, et al., 2007) but as reported by Wolf, they are typically built with specific applications in mind - such as a time-domain instrument for quantitative assessment of neonatal cerebral oxygenation monitoring. Thus, these systems tend to limit the flexibility that a custom device could achieve for BCI such as access to the unprocessed light levels from the detector and being able to develop advanced modulation schemes which may improve signal-to-noise ratios, which is paramount for highly attenuating tissue spectroscopy (i.e. seven to ten orders). Furthermore, most of these instruments are based on laser diodes and pose a potential risk (e.g. retinal damage) due to coherency effects. As will be discussed in Chapter 5, LEDs are sufficient for use as a light source in CWNIRS since the short transport scattering length of NIR light in tissue means any beam (e.g. from a laser diode) effectively becomes diffuse after the first few millimeters of penetration. Moreover, cerebral tissue of interest is 2-4 cm
below the scalp surface. Thus, for long-term BCI applications with unlimited exposure an LED-based solution is more satisfactory especially for vulnerable immobile populations such as ALS subjects. Furthermore, LEDs have a longer lifespan (~80,000 hours), and are more straightforward to modulate.
A multichannel implementation of an OBCI is hypothesised to increase the potential performance over what a single channel device can obtain. That is:
• Since more spatial coverage of the cerebral cortex (from the scalp probes) is obtained by increasing the number of channels, potentially more features are expected to be available to the BCI researcher that could be used for control of the device. Furthermore, visual, auditory, motor, and cognitive features have been measured using FNIRS and some of these could potentially be used for volitional control of an augmentative device. (However, most researchers have considered grand averages of these to examine the integrated effect of multiple instances of some task, e.g. finger flexion. Thus, a single-trial assessment of some of these features are explored in Chapters 6 and 7);
• With increased spatial coverage comes the possibility of localisation of regions of maximal activity to a mental event. These can only be reliably explored with a multichannel device, and thus the multichannel device developed in this dissertation facilitates those investigations (see Chapters 6 and 7);
• A multichannel instrument could be used for effectively overcoming anatomical differences between subjects, which could introduce inter-subject variability in BCI performance;
• Lateralisation of events for BCI is important and could be assessed. Since multi- regional functional activation to tasks, e.g. motor events stimulating cognitive areas or vice versa, or indeed bilateral activation from motor imagery (Coyle, 2005; Wriessnegger, et al., 2008) may increase classification difficulties, additional channels may be useful to compare signal strength from one hemisphere to the other. This is explored in Chapter 6 and 7 for mostly overt motor tasks;
• Additional channels also allow the researcher to develop noise cancellation techniques, especially immunity to global haemodynamic effects. The BCI group at NUI Maynooth is currently working on adaptive cancellation techniques to remove physiological noise and other artefacts (Matthews, et al., 2008a) based on the provision of multiple channels and instrumentation from this currently reported dissertation;
• A custom-made multichannel device also allows for exploring ways to reduce set-up time and perhaps deliver more computational time to fewer channels after an initial calibration. In other words, if multiple channels can identify similar localised functional activity, it may be possible to reduce the redundancy and remove some of those channels - if they are not necessary, e.g. they may be used for noise cancellation. This reduces the equipment placed on the subject and perhaps allows for more advanced online signal processing (although not reported in this dissertation) since more time could be allocated if fewer channels were being analysed;
• More channels assessing more functional cerebral events may allow for improvements in the potential data rate of the device from that of a single- channel device (6 bits/min). Furthermore, one cerebral region of tissue which is activated could be allowed to rest (return to baseline) while another basal-state region of tissue essentially takes over straight away, thus reducing the time needed for each event to be identified.
In Section 4.2, the chronological publication record of optical BCI was reviewed. In its 8 years, a custom-made dual-channel device was developed by Coyle (Coyle, et al., 2004; Coyle, 2005; Coyle, et al., 2007). Most others implemented signal processing techniques operate on the back of a commercial NIRS instrument, except for Bauernfeind (Bauernfeind, et al., 2008) whose essentially developed a single-channel version of the design from Coyle to compare against a commercial multichannel instrument. However, the design from Coyle was never intended to be a unit to be extended (i.e. modular), using multiple lock-in detectors etc. For BCI, a more flexible scheme was and is envisaged and is developed in this dissertation. Flexible in the sense that firstly, demodulation was moved completely into software using high-end National Instruments acquisition and generation cards and a data management scheme to ensure data buffering and triggering in a PCI eXtensions for Instrumentation (PXI) model. Moreover, for improved spatial specificity, multiple wavelengths were integrated into single high-power LED packages, custom designed and produced by Opto. Diode Corp. Thus, the proposal and methodology for the various design features of the versatile multichannel OBCI instrument are described next.