20 chronic stroke survivors (≥ 6 months from stroke onset) were recruited from the Queen Square Upper Limb rehabilitation programme (QSUL) for the experiment. The inclusion criteria for this experiment were: (1) first ever stroke and (2) no other brain injury, neurological condition or major psychiatric illness, while exclusion criteria were: (1) hemi-spatial neglect or hemianopia, (2) severe aphasia, or (3) pain limiting ability to participate in tasks or follow the study protocol. All participants were comprehensively informed about the study, and written, informed consent was obtained before their participation. Consent was obtained for all forms of personally identifiable data including clinical and kinematic data. The study was performed in accordance with the Declaration of Helsinki and was approved by the UCL/UCLH Joint Research Office (UCL Project ID Number: 17/0209; IRAS Project ID Number: 222832). The study was supported by a grant from the Jon Moulton Trust Charity.

The experiment was run using the KINARM Exoskeleton (BKIN Technologies Ltd, Kingston, ON, Canada); a robot which gathers kinematic data from the upper limbs during task performance. The KINARM Exoskeleton is a robotic device that supports the user’s limbs, forearms and hands, allowing only for horizontal movements involving shoulder and elbow joint flexion and extension (Fig. 1a). Participants sit with their limbs extended outward in a horizontal plane, typically at an angle of 85–90 degrees to the shoulder. The device includes an exoskeleton for each limb segment, customized for comfort and support. It also features a 2-D virtual reality display in the same plane as the limbs, providing visual targets and feedback. A calibration process before each session ensures accurate tracking and interaction within the virtual environment. Importantly, the robot provides gravitational support but does not assist in task completion in the experiment [22,23,24].

Prior to the start of the experiment consent was obtained from all participants. Participants underwent a one-time calibration process to accurately capture 2D movements in space. The experiment consisted of 2 stages: Stage 1 involved Practise and Normalisation (of motor ability), while Stage 2 included the Confidence Experiment (Fig. 1b) which were completed with both the more and less affected limb. The order was counterbalanced across participants. Each part was verbally explained thoroughly to the participants before the start of each experimental phase. In total, the experiment lasted for approximately 35–40 min which included breaks between experimental parts and when moving from one limb to the other.

In both Practice and Normalisation participants were asked to make reaching movements ‘as fast and accurately as possible’ to move a cursor (0.4 cm [2]) from a start target (2 cm [2]) to a peripheral target (1 cm [2]). The start target was positioned at the midpoint of each limb’s workspace (90° elbow flexion and 30° shoulder flexion), while the peripheral target was displaced 10 cm vertically along a straight line from the start target (Fig. 1c). The peripheral target appeared 300ms after participants moved into the start target. The start target turned green after a further 500ms which served as a Go cue. Participants started a new trial by moving their cursor back into the start target.

Practice – Provision of visual feedback of both cursor and target (vision ON).

Practice was included to familiarise participants with the task. Here participants had unlimited time to reach the peripheral target, which would disappear 300ms after it being reached. More specifically, participants were instructed to move their cursor ‘as fast and accurately as possible’ to the peripheral target. Importantly, constant visual feedback of both the cursor and targets was provided during each trial. In total participants completed 20 trials (each limb).

Normalisation of motor ability – No visual feedback of both cursor and target (vision OFF).

Normalisation was identical to the Practice phase except that visual feedback of both the cursor and the target disappeared once participants exited the start target. Participants were instructed to move their cursor ‘as fast and as close as possible’ to the peripheral target ‘without correcting your position after stopping’ (i.e., a single ballistic movement). Trials ended 2000ms after presentation of the Go cue. Importantly, participants did not receive any performance-based feedback (i.e., no binary feedback about task success and no visual feedback about target distance by showing the peripheral target again). Participants were asked to complete 20 no vision trials, which were used to normalise task success across participants and across more and less affected limbs as follows (Fig. 1d):

1) Determine reaching variability: The reaching movement end point (i.e., once reaching speeds fell below 2.0 cm/s) of each trial was determined (Fig. 1d (1)). Subsequently, a confidence ellipse with a confidence interval of 99.9% was calculated for the end point scatter to capture total reaching variability (Fig. 1d (2)).

2) Calculate target with 100% success rate: The area of the confidence ellipse represents the size of the peripheral target that would have yielded a ~ 100% success rate.

3) The width and height of this confidence ellipse was determined to create a new target with a predicted 100% success rate (T100) for both the more and less affected limb (Fig. 1d (3)). This procedure normalises reaching success across limbs for each participant.

4) Four further targets with success likelihoods of 65% (T65), 35% (T35), 15% (T15), and 5% (T5) were created (Fig. 1d (4)). To this end, 65%, 35%, 15%, and 5% of the height and length of T100 were calculated to create targets that have a 65%, 35%, 15%, and 5% success likelihood based on the original reaching end point scatter.

Note here that ‘confidence ellipse’ refers to a technique to determine the size of a scatter of data points with a pre-specified level of confidence. The confidence ellipse with a confidence interval of ~ 100% was used here to determine endpoint variability during reaching and is not related to participants confidence about task success.

Experimental design

a) Illustration of the KINARM Exoskeleton, which is a robot which gathers kinematic data from the upper limbs during task performance.

b) Experimental stages. Stage 1 involves Practise and Normalisation (of upper limb motor ability – see 1d), while Stage 2 involves the Confidence experiment (see 1e).

c) Illustration of the workspace and of the limb configuration. Participants were asked to make 10cm reaching movements from a start target to a peripheral target. Both were aligned so that the start target was at the midpoint of each limb’s workspace (90° elbow flexion and 30° shoulder flexion).

d) Illustration of the Normalisation procedure. During the second part of Stage 2 participants made 20 reaching movements but without visual feedback of both the cursor and the peripheral target once they exited the start target. Reaching ability was normalise using the following procedure: (1) Determine reaching end points of each trial, (2) build a confidence ellipse with 99.9% confidence interval around the scatter, (3) Obtain the target that has ~100% success likelihood (T100; i.e., participants would have hit this target in almost every trial), (4) Create targets based on T100 with increasing difficulty level (T65 – T5).

e) Illustration of a trial during the Confidence Experiment. Participants were presented with one of the 5 targets determined during Normalisation. (1) Prior to reaching participants were asked to rate how confident they were in hitting the target (Confidence Rating 1 – CR1), (2) After logging the verbal report participants aimed for the target without visual feedback of both the cursor and target, (3) After 2000ms participants were asked to rate how confident they were that they hit the target (Confidence Rating 2 – CR2), (4) After logging the verbal report, visual feedback was turned on again and participants could start a new trial by moving the cursor into the start target.

Confidence experiment – No visual feedback of either cursor and target (vision OFF).

We then investigated whether confidence about task success (estimated upper limb motor ability) matches actual task success (actual upper limb motor ability). Participants were not explicitly informed that their actual motor ability was normalised. Instead, participants were again instructed to move their cursor ‘as fast and as close as possible’ to the peripheral target ‘without correcting your position once having stopped the movement’, but this time without visual feedback of both the cursor and the target. (T100, T65, T 35, T15 or T5, Fig. 1d (4)). Specifically, participants aimed for the target, but once they initiated the movement (existed the target), the visual feedback of both the cursor and the target was turned off (vision OFF trials). Prior to each trial, participants were asked ‘How confident are you that you will hit the target?’ and responded (unlimited time) using a Likert scale (ranging from 1 to 5, with 5 indicating full confidence) in the centre of the workspace (Fig. 1e (1)). The Go cue was displayed 500ms after the confidence response and participants then performed the trial (with no visual feedback of the cursor or target (Fig. 1e (2)). The trial ended 2000ms after the go cue, when participants were asked ‘How confident are you that you hit the target?’ (CR2, Fig. 1e (3)) using the same Likert scale (ranging from 1 to 5) in the centre of the workspace (unlimited time). Participants completed 10 blocks of 5 trials (50 trials in total). Each block contained all targets (i.e., T100, T65, T35, T15, T5) which were presented in a random order. Therefore, participants completed 10 trials in each condition with each arm.

The 2D (x, y) position of the index finger was recorded at 1000 Hz by the KINARM Exoskeleton and was analysed ‘offline’ using Matlab (version R2019b, The MathWorks, Natick, MA, USA).

Confidence Rating (CR1 and CR2). Confidence ratings ranged from 1 to 5, with 5 indicating full confidence, in both CR1 and CR2. Medians were calculated because one-sample Kolmogorov-Smirnov tests (Matlab function kstest) indicated that the data were not normally distributed.

Success Rate (SR). Trial-based task success was operationalised as participants successfully hitting the target. Using the acquired 2D data (x, y) and the MATLAB function inpolygon we assessed if the participant hit the target at some point during the trial. SR represented the percentage of successful target hits.

1) Is confidence in achieving task success significantly different between the less and more affected upper limbs?

We performed a two-way repeated-measures ANOVA with ‘Arm’ (more and less affected limb) and ‘Task Difficulty’ (T100, T65, T35, T15, T5) as within factors and CR1 as the dependent variable. Post-hoc analysis included independent Wilcoxon Rank Sum Tests which were corrected for multiple comparisons using Bonferroni corrections while Cohen’s d was used to estimate effects sizes.

Additionally, we performed the same two-way repeated-measures ANOVA analysis as above using SR as the dependent variable, to assess differences in task success across limbs.

2) How is confidence about upper limb task success related to actual task performance?We performed a correlation between CR1 (median across all trials) and SR (% of successful target hits across all trials) independently for each limb.

3) Is confidence about task success and actual performance related to clinical measures including upper limb impairment?

To assess if confidence about task success or actual task success are a function of performance in clinical assessments, we ran independent correlations between both SR and CR1 and FMA scores. Furthermore, we ran the same analysis for FMA Sensory scores, HADS Anxiety sub scores, HADS Depression sub scores and NFI scores. Additionally, we performed a partial correlation between CR1 scores (prior reaching movement) to CR2 scores (post reaching movement) accounting for ‘Arm’ to assess if participants used post reaching feedback to update their confidence ratings.