A. Instrumentation
MMG was measured by a microphone-based sensor manufactured according to the method of Silva et al. [29]. A program was written in LabView to perform real-time data acquisition, contraction detection and switch activation. Microphone-detected MMG signals were continuously sampled at 1 KHz (NI USB-6210, National Instruments). The LabView program allowed online modification of parameters such as switch debounce time and activation thresholds, and provided the user with visual and auditory feedback when a muscle contraction was detected. On detecting a contraction, the DTR pin on a serial port of the computer was asserted. The serial port was interfaced with a conventional 1/8" mono-plug via an opto-isolator (4N36, Motorola Inc) to provide a standard switch output. A keyboard interface (KE-USB36, Hagstrom Electronics) was used with the mono-plug for computer access.
B. Contraction detection algorithm
Microphone signals were band-pass filtered with a 5th order Butterworth filter with a cut-off frequency range of 5-100 Hz. The low cut-off attenuates the effects of movement [30], while the high cut-off attenuates any noise beyond the accepted MMG signal range.
The contraction detection algorithm used in this study is a modification of the off-line activity-detection algorithm proposed by Alves and Chau [31]. In this study, continuous-wavelet-transform (CWT) coefficients of the MMG signal are compared to scale-specific thresholds to identify voluntary muscle activity of the frontalis muscle during small eyebrow raises. The CWT is defined as
where x
mmg
is the filtered MMG signal, and ψ is a mother wavelet shifted by k and scaled by a (k, a ∈ ℜ).
In the contraction-detection scheme, CWT transform coefficients at 14 scales, a, were compared to scale-specific thresholds, h(a), derived from baseline recordings. A muscle contraction event, z, is detected at sample k when the coefficients of at least j scales exceed their thresholds, i.e.
and
where K
baseline
are the samples corresponding to the baseline MMG signals and γ is the threshold-scaling factor.
The scaling-factor γ could be varied between 1.2 and 2.5 in increments of 0.2. The value of j was set to 1. CWT analysis was performed on 100 ms long MMG signals, using the sym7 mother wavelet at scales with pseudo-frequencies that spanned the 5-100 Hz frequency range of interest, i.e. a ∈ {7,9,10,12,14,15,17,20,23,28,35,46,69,115}.
C. Post processing, noise detection and switch debouncing
Figure 1 shows the procedure for converting the continuously acquired microphone signal, x, to a switch activation signal. CWT analysis was performed on the MMG signal, x
mmg
, using non-overlapping sliding windows, 100 ms in length. The output of CWT analysis is a muscle activity event, z [k], for each sample, k, of the windowed MMG signal. To reduce the probability of spurious activity being detected as voluntary contractions, when fewer than 10 ms of activity was detected in the 100 ms window, the activity was not considered a valid muscle event, i.e.
where m is the current window, and K = 100 is the window size.
CWT coefficients of MMG signals during eyebrow movement exceed those of artefact such as eyeblink and head movement. However, high-amplitude artefacts are observed in the MMG signal when the sensor is being moved during activities such as donning, doffing or adjusting the sensor position. While both contractions and movement are detected in the microphone signal associated with MMG (5-100 Hz), movement is more prominent and differentiable in the high-frequency microphone signal (100-300 Hz). Figure 2 shows an example of the low-frequency (MMG) and high-frequency components of the microphone signal during muscle contraction and sensor movement. The RMS of the high-frequency signal, xhf, shows good separation during contraction and sensor movement, and was therefore used to detect noise, n, at each window m of length K = 100 samples, i.e.
where threshold τ is determined from the maximum RMS of xhf during contraction. The noise event indicator was asserted if noise was detected in any of the M preceding windows, i.e.
In this implementation M was set to 5, thus disabling the switch if noise was detected in the preceding 500 ms.
The switch was enabled when a muscle event was detected and a noise event was absent. To avoid single contractions that typically last longer than 100 ms from being converted to multiple switch activations, the switch output was debounced with an adjustable delay. The delay was dependent on the speed at which the user could comfortably raise their eyebrow, and could be adjusted between 100-600 ms in 100 ms increments.
D. Events included in the baseline signal
The performance of the detection algorithm is profoundly affected by the choice of thresholds, and hence, the baseline signal that encompasses the artefact expected during switch use. Even when the forehead is at rest, the MMG signal recorded at the frontalis muscle is affected by visually-observable periodic artefact due to blood flow. As seen in Figure 3, the signal is further compromised by artefact due to eye-blinks and head movement. The characteristic MMG signal when the eyebrow is raised is an oscillatory wave whose amplitude initially rises and then decays. While the high amplitude at the initial burst of activity facilitates the detection of contraction onset, the eventual decay in activity encumbers activity-detection during sustained contractions. This limits the potential of a secondary switch activated by sustained eyebrow raises.
Figure 4 shows the maximum coefficients of the MMG signal during events such as rest, eye-blink, head movement, quick eyebrow raises and sustained frontalis contractions. The scale-specific thresholds of the detection algorithm are derived from the maximum coefficient of baseline MMG signals at each scale. The baseline includes MMG recorded during rest, blink and head movement. A contraction is detected if the CWT coefficient of at least one scale exceeds its baseline-derived threshold. The coefficients of the steady-state MMG during sustained contractions, while higher than the coefficients during rest, are confounded by those during movement artefact; therefore, sustained muscle activity cannot be detected. The signal transient at the initiation of contraction, however, has sufficiently high CWT coefficients to facilitate contraction-detection even during low-effort eyebrow raises. A quick and small contraction was therefore chosen as the preferred method for switch activation.
The detection algorithm was evaluated in real-time to monitor voluntary activity of the frontalis muscle and to generate a switch output.
E. Protocol for performance testing
A convenience sample of ten able-bodied individuals (5 male), age 27 ± 2 years, provided written consent to participate in the study. These participants, referred to as A1-A10 in this study, had no previous history of musculoskeletal illness. An adult with C1-C2 incomplete spinal cord injury (SCI), referred to as B1, was also recruited. B1's method of access included a sip-and-puff switch for wheelchair control, a head tracker (TrackerPro®, Madentec) for computer mouse emulation, and the dwell function (250 ms) of the head tracker for emulation of a mouse click.
Participants were instrumented with an MMG sensor [29] attached to the frontal belly of the occipitofrontalis muscle of the forehead with an elastic strap, as shown in Figure 5. The sensor was placed 1 cm above the eyebrow, above the inside corner of the right eye. Once the sensor was affixed, participants performed 30 s of 'baseline' activities such as blinking, talking, smiling and moving their head. Scale-specific thresholds were automatically evaluated from the baseline MMG signals using the contraction-detection software written in LabView. The threshold scaling factor was selectable in the 1.2-2.5 range, and was adjusted for each participant such that false activations due to blinks and movement were avoided and participants were able to activate the switch by raising their eyebrows with minimal effort. Once participants demonstrated that they could perform 10 consecutive cued switch activations correctly, the threshold parameters were set and remained unchanged for the remainder of the experiment.
Custom switch assessment software was written in Visual Basic to present participants with audio-visual stimuli and to record the times of switch activation and stimulus presentation. Participants were presented with a pseudo-random sequence of numbers at 2 s intervals, and were asked to activate the switch by raising their eyebrows slightly when the number "1" was presented. Participants performed four trials of the experiment, with a 30 s break in between trials. One-hundred stimuli were presented during each trial, with the actionable stimulus (i.e. number 1) being presented 25% of the time. Throughout the session, participants were encouraged not to sit absolutely still, but rather to behave in a manner that they normally would when seated at a desk: they were free to blink, sway their chair slightly, move their head and talk without moving their eyebrows or the strap. The number of true positives (TP), true negatives (TN), false positives (FP) and false negatives (FN) were recorded during the cued stimulus tests.
In addition to responding to cued stimuli, participant B1 typed a pangram for each of two selection modalities: dwell and eyebrow-raise. For both typing tasks, B1 used the head-tracker to point to a character on an on-screen keyboard. For the first task, B1 dwelled at the character's location for 250 ms to select it; this was the method B1 regularly used for typing for more than seven years. For the second task, B1 raised his eyebrow to select the character. The time taken to complete each task was recorded.
After the data-collection trials were completed, all participants practiced using the switch for 1 hour, performing activities such as typing using a scanning keyboard. At the end of the hour, participants were asked to rate the level of effort and fatigue associated with controlling the eyebrow switch on a five-point linear scale: [1-Nothing at all, not tired; 2- A little, not tired; 3- Moderate, a little tired; 4- A lot, tired; 5-Too much, very tired]. In addition, participants were asked to rate if they had to try multiple times before activating the switch: [1-Never; 2- Very infrequently; 3- Sometimes; 4- Very often; 5- Almost all the time].
The experimental protocol was approved by the hospital and university research ethics boards, and was in compliance with the Declaration of Helsinki.
F. Performance Metrics
The sensitivity and specificity of the MMG switch were evaluated from the cued stimulus test, and are given by
and
Sensitivity is a measure of correctly identified muscle contractions, while specificity is a measure of correctly rejected artefacts.
Trends in response delay were used to gauge if participants were fatigued from prolonged use of the eyebrow switch. For each participant, a linear regression of response delay against elapsed session time was evaluated, and the 95% confidence-interval (CI) of the slope was computed. Here it is assumed that response time increases with increasing fatigue.