1. ii
ABSTRACT
Motivation: Conventional surface electromyography (EMG) methods cannot be used to detect deep
muscle activation. A new non-invasive superficial and deep muscle EMG (sdEMG) technique has
recently been used to derive the EMG activity of Brachialis and Tibialis Posterior muscles in the
upper and lower limb respectively. The aim of the present study was to apply a modified version of
sdEMG to the forearm to detect EMG activity of the deep extrinsic thumb muscles Flexor Pollicis
Longus (FPL), Extensor Pollicis Longus (EPL), Extensor Pollicis Brevis (EPB) and Abductor Pollicis
Longus (APL) using surface electrodes.
Methods: High density monopolar EMG was detected from 2 concentric rings, each consisting of 20
custom designed and manufactured silver electrodes, placed at the distal and proximal thirds of the
right forearm of 15 healthy male participants. The EMG signals were recorded by a custom
synthesised from open source components, EMG amplifier system interfacing with a custom
designed LabVIEW® program. The participants performed 10 repetitions of isometric thumb flexion
(TFl), thumb extension (TEx), thumb abduction (TAb), thumb adduction (TAd), index finger flexion
(IFFl) and index finger extension (IFEx). Each isometric contraction was performed in a randomized
order at a standardized effort level of 30% of the participant’s maximum voluntary contraction
(verified by a custom designed and built thumb dynamometer). The Independent Component
Analysis (ICA) algorithm, fastICA, was used to un-mix the 40 monopolar EMG waveforms (containing
EMG activity attributable to both superficial and deep muscles) into 40 constitutive components,
known as the Independent Components (ICs). The activation envelope of the ICs was found using a
250ms RMS smoothing filter and normalized between 0 and 1. A contraction sequence specific
predicted EMG waveform based on intramuscular measurements (from existing studies in the
literature) was created for each deep muscle and correlated with the processed ICs using Pearson’s
Correlation Coefficient (r). The ICs were ranked according to the corresponding r value and the
highest r ranked IC for each muscle was considered to represent the recovered EMG activity from
that particular muscle. Finally, a per sample basis accuracy, sensitivity and specificity analysis was
conducted between each deep muscle’s predicted EMG and highest r ranked IC at different
activation thresholds. A linear mixed-effects statistical model was used to find the overall accuracy,
sensitivity and specificity values over all the thresholds for each deep muscle.
Results: Overall correlations of 0.81 for FPL (D), 0.88 for EPL (D), 0.92 for EPB (D) and 0.83 for APL (D)
(p<0.001 for all muscles) were found between the predicted EMG waveforms and ICs. Using an
activation threshold of 3 standard deviations above a resting baseline level, statistically significant
(p<0.001) accuracy, sensitivity and specificity measures were found between the predicted EMG
waveforms and top r ranked ICs for each of the deep muscles. The values of the 3 statistical
measures (accuracy, sensitivity, specificity) for each of the deep muscles were: FPL (0.76, 0.88, 0.70);
EPL (0.87, 0.85, 0.91); EPB (0.94, 0.93, 0.94); APL (0.80, 0.87, 0.87).
Conclusions: The results indicate that this is the first non-invasive detection of the EMG activity of
FPL (D), EPL (D), EPB (D) and APL (D). The ability to detect movement intention as a result of
activation from these muscles may be of use for robot based targeted rehabilitation of the hand or
in the control of prosthetic hand devices.