2.6.2. TMS data

The data files were extracted from Signal into MatLab and were analysed offline using custom-written routines in MatLab (2018a, Mathworks). Two dependent variables were measured on a trial-by-trial basis as follows: (1) MEP peak-to-peak amplitude (Fig. 1C) and (2) reaction time (RT) measured from the time of the IS to the onset of volitional muscle activity (Fig. 1C). Peak-to-peak MEP amplitudes for each condition were estimated from the acquired EMG signal without applying any additional filters. A logarithmic transformation (to the base of e) of single-trial MEP amplitudes was performed before the statistical tests to ensure normality of the samples. Resting EMG was defined as the root mean square (rms) across all trials for each participant in the first 100 ms of each trial (prior to the WS). Thresholds set at five times these levels were used to determine the RT. All trials were then visually inspected and manually corrected to ensure that RT was estimated properly, there was no undue influence of the silent period following stimulation and that no build-up of EMG was apparent before the TMS. Trials in which RT was less than 75 ms or greater than 500 ms were excluded from the final analysis as they represented premature and late responses respectively. Trials were also excluded if the MEP amplitude was less than 0.025 mV. Trials containing outlier MEP amplitudes (Grubb’s test, p < 0.005) were also excluded from the final analysis. On average, 15.4% of TMS trials and 16% of catch trials were excluded across all stroke survivors with a minimum of 7 trials per stimulation condition.

To examine the effect of fatigue on corticospinal excitability and RT, log-transformed MEP amplitudes and RTs were labelled according to the time at which TMS was delivered (WS, WP, IS, RT30, RT50, RT70) and analysed by means of generalized mixed effects models carried out within the R environment for statistical computing (RStudio Version 1.2.5033), using the ‘lme4′ package (Bates et al. 2014). The ‘lmerTest’ package (Kuznetsova et al. 2017) was used to estimate the p-values for the t-test based on the Satterthwaite approximation for degrees of freedom. A stepwise ANOVA based on Satterthwaite’s approximation of degrees of freedom for model selection (lowest AIC value and p-value) was used to identify the combinations of variables that best predicted the outcome variables (MEP amplitude and RT). Based on previous studies we had reason to believe that the change in MEP amplitude over time would follow a quadratic trend whereas RT would follow a linear trend (Bestmann and Duque, 2016, Hannah et al., 2018, Ibáñez et al., 2020). Therefore, we compared the AIC for both the linear and quadratic fit for both MEP amplitude and RT (Table 2). The quadratic model was a better fit for the MEP amplitude data while a linear model was a better fit for the RT data and the effect of Warning. A similar analysis was used to examine the effect of fatigue and condition (Warning condition vs No Warning condition) on corticospinal excitability and RT (Table 2). Assumptions of normality and homoscedasticity of the residuals for each model were assessed visually using quantile-quantile normal plots and fitted- versus residual-value plots. Individual spearman’s rank correlations were carried out between FSS-7 and the dependent variable in each model.

The result of generalized mixed effects model comparisons across the corticospinal excitability data, the reaction time data and the effect of warning on both corticospinal excitability and reaction time. Participants nested in time are the random effect in each model. Significance levels are indicated by * (* < 0.05). Df = degrees of freedom; AIC = Akaike’s information criterion; Chisq = chi-squared statistic; Chi Df = chi-squared degree of freedom; Pr(>Chisq) = probability value.

For a graphical representation of the results, stroke survivors were divided into two groups, high and low fatigue, based on their FSS-7 scores. A cut-off score of less than four on the FSS-7 was classified as low-fatigue and a score equal or greater than four was classified as high fatigue (Valko et al. 2008). Throughout the analysis, FSS-7 was treated as a continuous scale.