Table 1.

Summary of studies on electrically evoked contraction and evoked electromyographic (EMG)-force relationship.

Reference	Clinical Population and Study Design	Parameter Analysed	Outcome/Findings
Mizrahi et al., 1997 [17]	1-T4 and 1-T6/7 SCI individuals Transcutaneous isometric stimulation of quadriceps muscle with PW of 0.25 ms SF of 20 Hz 16 weeks of 45 min stimulation per week Level of Intramuscular pH was used to represent fatigue within the contractile element of the muscle model Measured at primary and post-recovery fatigue stage.	PTP, RTP, AVREC, RMS, TSP, MDF	Force-eEMG relationship was correlated by PTP and RMS (r = 0.97 and r = 0.95, p < 0.05) in fatigue and post fatigue period, respectively. Force-eEMG parameters showed that metabolic and electrolytic factor may be significant in assessing recovery and fatigue. IM pH decreased to 6.2 and correlated to the decay in the stimulated quadriceps force.
Erfanian et al., 1998 [43]	2 complete T7 SCI individuals VL was activated under isometric condition at one joint angle Artefact balancing was used to remove stimulation artefact 6 different percutaneous stimulation patterns were adopted Constant SF and amplitude of 20 Hz and 20 mA was used, respectively.	MAV of eEMG	Evoked EMG predicted muscle torque at only one angle. Mean square error (MSE) of 0.0383 was obtained as performance index and showed the quality of prediction. The approach is only viable for intramuscular stimulation. No verification yet on multi-muscle, i.e., FES practical use.
Ding et al., (2005) [44]	14 SCI (all except one has thoracic level motor compete lesion). Transcutaneous stimulation was used to evoke isometric force of the quadriceps femoris muscle. Each subject participated in pre-fatigue (1 stimulation train every 20 s) and after 10 m rest fatigue protocol (110, 13-pulse, 40 Hz trains, i.e., fatigue inducing train). Stimulation trains were delivered with a 700 ms rest time between successful trains.	PF, FTI	The predictive model was recommended for FES application because of the rapid parameter identification, fast optimization analysis and accurate prediction for feedforward control. The ICCs between the experimental and predicted force-time integrals and peak forces were above r = 0.90, p = 0.05. However, the model could only predict the force response of quadriceps at one length, under isometric contraction at only one knee joint angle. Prediction of muscle force for real time FES functional activities was only recommended.
Zhang et al., 2011 [45]	5 SCI individuals (3-T6, 1-C5, 1-C7) Right triceps surae muscle group was activated to generate isometric ankle torque with constant SF of 30 Hz and PW of 0.45 ms Surface stimulation was adopted to plantarflex the ankle joint Measurement taken during: fatigue inducing test, fatigue recovery test and random test	MAV of eEMG	Torque prediction model based on Hammerstein structure properly fitted muscle model under isometric condition. With 18 s prediction horizon, RMS and Peak prediction errors were 0.097 and 0.34 maximum, respectively. Dynamic muscle action is necessary to validate the model for FES practical application. Reliability of the model was not investigated.
Hayashibe et al., 2011 [10]	1 complete T8 SCI individual Quadriceps and tibialis anterior muscles were activated by an implanted neural stimulation of peroneal nerves with the fixed SF of 30 Hz and PW of 0.6 ms Stimulation strategy was chosen to induce high level of fatigue Dynamometer was used to measure the torque of isometric ankle dorsiflexion	MDF, MAV of M-wave	MDF of M-wave was correlated with the torque during fatigue (r = 0.77, p < 0.05) but not during potentiation Normalised RMS deviation was 0.145 and 0.00884 for the prediction indices. Torque prediction at low magnitude of EMG was less accurate. The study suggested significant relationship between M-wave and torque only if implanted stimulation is used.
Li et al., 2012 [6]	2 T6 level SCI individual used for validation of the muscle model Transcutaneous stimulated with the SF of 30 Hz and PW of 0.45 ms was used during isometric condition The relationship between eEMG and torque/force with Non-linear Arm type recurrent neural network (NARX-RNN) model was demonstrated.	MAV of eEMG, Muscle torque	For prediction horizons of (10, 50, 70 s) the RMS error ranges from 0.0402 to 0.1067 Model performance on muscle dynamic action was not verified.
Hwang et al., 2012 [46]	4 incomplete SCI and 4 Healthy Volunteers Transcutaneous stimulation of fixed SF of 20 Hz and PW of either 0.4 ms or 0.8 ms was adopted to activate quadriceps muscle group VEMG, eEMG and the combination of both were measured separately to determine the contribution of each to force/torque estimation RBF neural network algorithm adopted to assess torque using parameters derived from VEMG and eEMG	RMS of VEMG, M-wave of eEMG	RMS of VEMG was shown to estimate torque but the performance was poor during validation in the feedback control system. The processing speed was low for an on line control application.

Abbreviation: PW: Pulse width; SF: Stimulation frequency; VL: Vastus lateralis; PTP: Peak to peak amplitude; RTP: Rise time to peak amplitude; AVREC: Average rectified; RMS: Root mean square; TSP: Total spectra power; MDF: Median frequency; MAV: Mean absolute value; Peak torque: PTV; Force-time integral: FTI; EMG: Voluntary EMG; eEMG: Evoked EMG; RBF: Radial basis function; IM: Intramuscular.