Investigating the Intervention Parameters of Endogenous Paired Associative Stimulation (ePAS)

Advances in our understanding of neural plasticity have prompted the emergence of neuromodulatory interventions, which modulate corticomotor excitability (CME) and hold potential for accelerating stroke recovery. Endogenous paired associative stimulation (ePAS) involves the repeated pairing of a single pulse of peripheral electrical stimulation (PES) with endogenous movement-related cortical potentials (MRCPs), which are derived from electroencephalography. However, little is known about the optimal parameters for its delivery. A factorial design with repeated measures delivered four different versions of ePAS, in which PES intensities and movement type were manipulated. Linear mixed models were employed to assess interaction effects between PES intensity (suprathreshold (Hi) and motor threshold (Lo)) and movement type (Voluntary and Imagined) on CME. ePAS interventions significantly increased CME compared to control interventions, except in the case of Lo-Voluntary ePAS. There was an overall main effect for the Hi-Voluntary ePAS intervention immediately post-intervention (p = 0.002), with a sub-additive interaction effect at 30 min’ post-intervention (p = 0.042). Hi-Imagined and Lo-Imagined ePAS significantly increased CME for 30 min post-intervention (p = 0.038 and p = 0.043 respectively). The effects of the two PES intensities were not significantly different. CME was significantly greater after performing imagined movements, compared to voluntary movements, with motor threshold PES (Lo) 15 min post-intervention (p = 0.012). This study supports previous research investigating Lo-Imagined ePAS and extends those findings by illustrating that ePAS interventions that deliver suprathreshold intensities during voluntary or imagined movements (Hi-Voluntary and Hi-Imagined) also increase CME. Importantly, our findings indicate that stimulation intensity and movement type interact in ePAS interventions. Factorial designs are an efficient way to explore the effects of manipulating the parameters of neuromodulatory interventions. Further research is required to ensure that these parameters are appropriately refined to maximise intervention efficacy for people with stroke and to support translation into clinical practice.


Supplementary Material
History of specific repetitive motor activity 0 0

Methodological factors
Position and contact of EMG electrodes 1 1 Amount of relaxation/contraction of target muscles 1 1 Prior motor activity of the muscles to be tested 1 1 Level of relaxation of muscles other than those being tested NA 0 Coil type (size and geometry) 1 1 Coil orientation 1 1 Direction of induced current in the brain 1 1 Coil location and stability (with or without neuronavigation system) 1 1 Type of stimulator used (e.g., brand) 1 1 Overall % Score 94% 1 = yes, 0 = no, NA = non-applicable. Statistical analysis plan 1. Analysis sets and datasets

Missingness at random assumption
It is assumed that missing data are missing at random (MAR). The planned linear mixed effects analysis will adequately account for missing data under this assumption [Carpenter 2007].

Datasets
There will be two datasets used in the analysis. One will be the absolute dataset, consisting of the absolute values of the outcomes pre-intervention, immediately post-intervention, and at 15, 30, and 45 minutes post-intervention (normally, 5 time-points). All measurements at all time-points will be retained. The second one will be called the relative dataset and will consist of the relative change from the pre-intervention measure immediately post-intervention and at 15, 30, and 45 minutes postintervention. These relative changes will be computed from the mean of the measurements at each time-point and a weight corresponding to the number of observations retained post-intervention to compute the mean at each post-intervention time point. The relative change was calculated as follows (post-pre)/pre × 100.

Study outcomes and baseline covariates
2.1. Cortico-motor excitability (CME) outcomes CME outcomes are measured pre-intervention and at 0, 15, 30, and 45 minutes post-intervention. They consist of:

Baseline covariates
The covariates collected pre-intervention consist of maximum voluntary contraction (MVC) and active motor threshold percentage (AMT). These covariates have been tested for adjustment in the models during a blind review (using the blinded treatment codes to adjust for treatment), using a 5% significance threshold to decide on inclusion. Pre-intervention MEP amplitude and MEP area values have also been tested in the same manner.

Notes on the data
We distinguish between Blind Intervention and the Blind Intervention Group in the absolute data only.
Blind Intervention takes on the values "None", "A", "B", "C", "D", "E", and "F", with "None" applying in all pre-intervention observations. It is used to define fixed effects.
The Blind Intervention Group takes on values "A", "B", "C", "D", "E", and "F" and refers to the intervention applied in a whole session, including the pre-intervention measurements. It is used to define random effects.
Only the Blind Intervention Group is used in the relative data, as the pre-intervention observations are not included in the analyses.

Inferential framework
The inferential framework selected is linear mixed modelling. The large size of the datasets (over 7300 observations in the absolute data and over 580 observations in the relative data) render concerns about non-normality of the secondary residuals, in spite of the dependence between the observations, if we extend the arguments of Lumley and colleagues [Lumley, Diehr, Emerson, and Chen 2002] regarding linear regression to linear mixed regression. Analyses were carried out using the package lme4 [Bates, Machler, Bolker and Walker 2014] in R (R Core Team 2018) and SAS/STAT™ software.

Blinded selection of model
During the blind review, the model to be used will be selected from amongst 144 models (absolute data) or 288 models (relative data) defined with the core elements listed below. A final assessment of residual covariance structure and heteroscedasticity across the Blind Intervention Groups in the retained model will be carried out.
Core covariate: -Blind Intervention: None, A, B, C, D, E, F (None is applied to the pre-intervention data) Core random effects: -Participant random intercept: P01,…,P025 -Participant and Blind Intervention Group interaction random intercept: P01*A, P01*B, etc. The model also involves the following alternative components. Alternative covariates: (1) Time-point: either time as a factor (PRE, POST0, POST15, POST30, POST45), time as continuous (PRE=0,…,POST45=4), or time as continuous as well as its square (2) Time-point and Blind Intervention interaction: either none, or in interaction with time as a factor, or in interaction with time as a continuous covariate (Note: For covariates only, we do not mix categorical and continuous in the same model.) (3) MVC: Present or not (4) AMT: Present or not (5) Baseline MEP value: Present or not (relative data only) (6) Order of the intervention in the cross-over: Present or not Alternative random effects (7) Participant random slope for time as continuous: Present or not (8) Participant and Blind Intervention Group interaction random slope for time as continuous: Present (only if 7. is present) or not Notes: Time was taken as continuous in random effect regardless of fixed effects. Model selection was automated up to the selection of items 1-8 with the use of the package lme4 [58] in R (R Core Team, 2018). Further covariance investigation was carried out using PROC MIXED, part of the SAS/STAT software. In case item 7 was retained, alternative formulations for the variance structure were investigated (R-side covariance structure), with time as a continuous index and instances of participant-interventions as the subject. These covariance structures were: -compound symmetry -heterogeneous compound symmetry -autoregressive of order 1 -autoregressive moving average of order (1, 1) -spatial exponential -spatial Gaussian -spatial power covariance. In all cases, heterogeneity of the variance parameters by treatment group was investigated, always using AIC as a criterion. Under failure to converge, the model was deemed inadequate without further investigation.

Absolute data
The final model was selected based on Akaike's information criterion and corresponded to the following model for both MEP amplitude and MEP area.

Retained fixed effects:
-Blind Intervention -Time as a factor -Blind Intervention in interaction with time -AMT -MVC Retained random effects: -Participant random intercept -Participant in interaction with Blind Intervention Group random intercept and random time slope Other variance parameters: -Variances were found to be heterogeneous across Blind Treatment Groups.

Relative data
The final model for the relative data was selected based on Akaike's information criterion and corresponded to the following model for both relative MEP amplitude and MEP area.
Retained fixed effects: -Blind Intervention -Absolute baseline value Retained random effects: -Participant random intercept -Autoregressive moving average of order 1,1, with continuous time as the index and participant-time as the subject Other variance parameters: -Variance parameters were found to be heterogeneous across Blind Treatment Groups.

Translation to unblinded model
The actual model retained for the primary analysis is the factorial version of the retained model. Movement type takes on the values "None", "Real", or "Imagined"; intensity takes on the values "NA", "0%", "100%", or "300%". The values "None" for movement type and "NA" for intensity only apply in the pre-intervention phase for Tx equal to "None".