In- and outpatients after a stroke were recruited in a neurorehabilitation clinic in the Northwestern part of Switzerland between September 2019 and July 2021. Due to the COVID-19 pandemic, recruiting had to be suspended from March to August 2020. Potential study patients received oral and written information and had at least 24 h to consider participation. Data collection began after signing the study consent form. Patients were included in the study if they fulfilled the eligibility criteria listed in Table 1.

The investigation was conducted in accordance with the Declaration of Helsinki and was approved by the local ethical committee (EKNZ no: 2019-00348, 2020-00545). Patients underwent two measurement sessions within 7 days at least 48 h apart.

Patients' MI ability was assessed using the short version of the Kinaesthetic and Visual Imagery Questionnaire (KVIQ-10), the body rotation task (BRT), and the mental chronometry (MC) task. These assessments evaluate distinct aspects of MI ability and are suggested to be applied in combination for an accurate assessment of individuals' MI ability (Di Rienzo et al., 2014). All of them are commonly applied for MI ability assessments in neurological patients (Heremans et al., 2013).

The KVIQ-10 is a reliable and valid instrument to evaluate the ability to visualize and feel the imagined movements in patients with sensorimotor impairments (Malouin et al., 2007; Schuster et al., 2012). Patients rate the clarity and sensation intensity of five imagined movements using a five-point rating scale: from 1 = no image to 5 = as clear as seeing, respectively, 1 = no sensation to 5 = as intense as executing the movement.

The BRT refers to an unconscious process, called implicit MI, and it determines individuals' ability to imagine moving their internal representation of their own body parts into a presented position or angle (Fiorio et al., 2006). Patients were required to determine the laterality of 32 pictures of a human hand or foot. The number of correct answers determined their score.

The MC examines the temporal coupling of MI (Malouin et al., 2008). To determine MC, the ratio of time needed to imagine and execute a grasping task was calculated.

To record eye movements on patients' affected body side, the EOG signals were recorded with a sampling frequency of 1000 Hz. After skin preparation, one pair of surface electrodes (BlueSensor N, Ambu, Germany) was placed on the vertical axis over the pupil. Data were captured using wireless sensors (Myon aktos classic, myon JSC, Switzerland) and the proEMG stand-alone software (V.2.1, Prophysics JSC, Switzerland). Simultaneously, patients wore the J!NS MEME smart eyeglasses (JIN CO., LTD, Japan, Figure 1). The smart eyeglasses are equipped with a 3-axis accelerometer, a 3-axis gyroscope sensor, and a 3-point electrooculography sensor integrated into the nose pads. The metal dry electrodes, two on both sides of the rhinion and one on the nasion, measured the electrical potential in μV, thus detecting horizontal and vertical eye movements and blinking frequency (Kanoh et al., 2015). Data acquisition took place with a sampling frequency of 100 Hz while the signal was transmitted in real time via Bluetooth connection using the J!NS MEME Data Logger software (V.1.1.10, JIN CO. LTD., Japan).

Cortical activity was measured using a 64-channel EEG cap (BE PLUS LTM Prewired Headcap, EBNeuro, Italy) and a wireless headbox with a sampling frequency of 1000 Hz (during the preprocessing, the frequency was down-sampled to 250 Hz). The position of the Ag/AgCl electrodes followed the 10–20 international system. Electrode impedances were kept below 20 kΩ, and data were acquired using Galileo NT Line software (2016, EBNeuro, Italy). For the ERD/ERS analysis, alpha (8–12 Hz) and beta (13–30 Hz) bands were evaluated.

Furthermore, oxygen saturation (SpO₂) and heart rate (HR) were monitored and recorded continuously using a pulse oximeter that was included in the EEG recording device.

Recordings from the different measurement systems were synchronized using an external data acquisition system (LabJack U3-LV, LabJack Corporation, Lakewood, USA).

Patients sat in front of a computer screen and received auditory and visual instructions for the tasks from the speakers of a connected laptop (Figures 2A, B). The task was designed using the PsychoPy software (3.0./2020.1.3., Open Science Tools Ltd., Great Britain) (Peirce et al., 2019). A customized chinrest (Gerald Kann—Kinnstützen und mechanische Vorrichtungen, Magdeburg, Germany) restricted head movements. The following standardized recording procedure was applied:

During the complete recording procedure, patients were asked to keep their eyes open and were instructed to begin the task after the auditory and visual cues. They were encouraged to observe their movement during PE of the motor tasks. Each run compromised 21 trials of MI and PE of the task in blocks of three (Kobelt et al., 2018). A PE block was constantly followed by an MI block. A trial consisted of 5-s instruction and 2- to 4-s PE or MI period. There were two runs, which led to 42 physical and 42 imagery repetitions of the task. The patients had a rest between the runs. Thus, the MI and the PE data collection lasted about 6 min each (9 s × 21 trials × 2 runs). The experimental paradigm is shown in Figure 3.

An event start point was labeled precisely according to the recording log-files: the time point of the end of auditory and visual cues. During the experiment, the assessor monitored the exact termination of the physical and imagery task: if the cup was placed back to the starting point or if the patient indicated the end of the imagination by a tap with his/her non-involved hand. The assessor marked these ends as event points in the recording log-files by pressing a keyboard key.

Based on the literature, 14 to 20 participants are sufficient to evaluate psychometric properties and to recognize MI-related eye movements (Heremans et al., 2012a,b; Jang et al., 2018). Additionally, we performed a sample size calculation based on our EOG and SE data measured during the previously conducted technical validation of the setup and experimental task with five healthy individuals. Using the G^*Power software (V.3.1, Heinrich Heine University Düsseldorf, Düsseldorf, Germany; Faul et al., 2007), a sample size of 40 was estimated with a power level of 0.8 and alpha level of 0.05 with an expected correlation r ≥ 0.60 between oculomotor activity during MI and PE. However, during the study implementation the COVID-19 pandemic forced us to reduce the number of included patients.

The correlation of smart eyeglasses signals between MI and PE varied between r = 0.12 and r = 0.76. The cross-correlation analyses of smart eyeglasses signals between MI and PE revealed mean correlation value r = 0.2 for task pA/iA, mean correlation value r = 0.16 for task pB/iB, and mean correlation value r = 0.19 for task pC/iC.

The cross-correlation was calculated for each patient and each condition (PE, MI). Table 3 presents the range of the correlation values for each task (PE: pA, pB, pC; MI: iA, iB, iC). The highest estimate revealed the correlation during task pB with the value r = 0.71 and during task iB with the value r = 0.67 that described vertical eye movements. The mean correlation during PE overall (pA, pB, pC) was r = 0.57 and during MI overall (iA, iB, iC) r = 0.53.

The ICCs were calculated for each patient, each condition (PE, MI), and each task (PE: pA, pB, pC; MI: iA, iB, iC). Every task revealed similar ICCs ranging between 0.51 and 0.54. The average ICC was 0.53 (95% CI 0.29–0.76) for PE and 0.51 (95% CI 0.26–0.80) for MI. All ICCs with confidence intervals are listed in Table 3.

The time-frequency maps for electrode positions C3, Cz, and C4 showed large variability during different events. In individual cases, similar ERD and ERS patterns could be recognized by visual inspection during MI and PE. Figures 5A, B visualizes an example of band power 8–12 Hz (A), band power 13–30 HZ (B) time course displaying ERD and ERS during MI and PE of the grasping task for one patient (P009) and for electrode location C3. Average ERD/ERS values were compared for every patient in 10 different electrodes for MI and PE events using mixed-effect ANOVA. There were no statistically significant similarities between the events (MI/PE) except for event C in the F4 electrode in the beta power band (p = 0.0373). Additionally, in the alpha band, time and age showed a significant effect for the CZ electrode. This indicates that the older the patient more the similar the ERD/ERS values during MI and PE. As the experiment progressed, the ERD/ERS values became more similar during the MI and PE conditions. ERD/ERS values used for ANOVA were the average values during all physical (pA, pB, pC) and imagery (iA, iB, iC) events. The ERD/ERS values were then examined for each event separately and are displayed in Table 4.

Descriptive statistics for the collected SpO₂ values showed a normal distribution when tested with the Shapiro–Wilk test. The data showed no outliers in the referring range (1.5^* interquartile range). The HR datasets for all conditions followed a normal distribution determined with p > 0.05 in the Shapiro–Wilk test. The statistical description of the SpO₂ and HR data collected for the upper limb task and the details for the distribution are presented in Table 5.

The results of the MI ability assessments were normally distributed as indicated by the Shapiro–Wilk test. The details of the descriptive analysis are shown in Table 5. The descriptive statistics (Table 5) for the different conditions showed no major differences in the SpO₂ values but for HR.

The results of the linear mixed model with the corresponding estimates and the confidence intervals are presented in Table 6. According to the linear mixed model, no differences in the SpO₂ values were identified between all experimental conditions. The statistical analysis of the applied linear mixed model for the HR values revealed a significant difference when the imagined condition was tested against the resting condition. Conversely, HR showed no significant difference from the imagined to the active condition. The complete analysis is presented in Table 6.

The results of the correlation analysis for the SpO₂ and HR are summarized in Table 7. No significant correlations were found between the changes in SpO₂ during MI and the scores of MI ability assessments. No significant correlations were found between the changes in HR and the scores of MI ability assessments.

Oculomotor activity measured with smart eyeglasses during MI demonstrated a low correlation with those during PE. Here, the recording frequency could have an impact on the results. Eye movements are usually recorded with a sampling frequency of 1024 Hz that allows to record fine movements (Heremans et al., 2008, 2009, 2011). The recording frequency of the SE in our study was predefined to 100 Hz only, which might not have been sufficient for recognizing MI. A further explanation could be the different strategies patients used to complete the task. In our experiment, it was not possible to directly control how precisely patients imagined the movement. Patients might have used a different strategy to complete the task, and eye movements were superfluous or they rather have used peripheral vision. To support this explanation, Heremans and colleagues also reported that 11 to 17 % of their healthy participants did not always show task-related eye movements (Heremans et al., 2008, 2009). Moreover, our patients used their self-selected speed to execute or imagine a movement, which could have resulted in an additional source of variability of eye movement patterns. This explanation is supported by the findings from Heremans et al. (2009), who stated that external auditory cues offered by a metronome enhance temporal accuracy.

However, our selected task itself could be taken into account. Studies investigating eye movements with EOG usually demand participants to carry out a strictly directed task. They apply visual or/and auditory cues (e.g., using a metronome to instruct the start of a movement and visualizing the start and end point of a task with a symbol), thus influencing duration and spatial accuracy (Heremans et al., 2008, 2012a,b; Lanata et al., 2020). In contrast, in our study, patients were free to choose how to perform tasks. We could conclude that the experimental task selected for an MI ability assessment requires more external control or cueing and standardization.

Additionally, studies describing large similarities in eye movement during MI and PE examined different parameters, which could explain the differences in our results. They analyzed eye movement amplitude (corresponding distance), number of eye movements, duration, or saccades (Spivey and Geng, 2001; Laeng et al., 2002; de'Sperati, 2003; Heremans et al., 2013) while we compared the shape of the EOG signal.

The validity of the smart eyeglasses was examined using a cross-correlation. The overall mean correlation value between the signals of the smart eyeglasses and conventional EOG during PE was r = 0.57 and during MI r = 0.53. According to the correlation classification for outcome measures in stroke rehabilitation of Salter et al. (2005), our findings imply an adequate cross-correlation. However, during physical task pB and imagery task iB the smart eyeglasses revealed a correlation value of r = 0.71 and r = 0.67 that fulfilled the desirable value r of ≥ 0.60 for excellent correlation (Salter et al., 2005). The task pB and iB involved mainly vertical eye movement. The higher values during these tasks can be explained by the electrode placement. Previous studies used four surface electrodes for conventional EOG (Heremans et al., 2008, 2009, 2011). In the current investigation, only one pair of electrodes was placed above and under the eye in a vertical line; consequently, more activity could be recorded when the eye was moving vertically. Another pair of electrodes were waived due to practical reasons. As EOG was simultaneously recorded with the smart eyeglasses and the surface electrodes, the smart eyeglasses occupied the nose and made the attachment of an additional pair of surface electrodes impossible. The head and face were already heavily covered with the smart eyeglasses and the EEG cap.

In addition, the recording frequency could have had an impact on the results. However, our conventional EOG was recorded with a sampling frequency of 1000 Hz. For comparison with the smart eyeglasses data EOG signals had to be down-sampled, which entails a great loss of accuracy thus affecting the correlation results. Yet, Barbara and Camilleri (2016) presented a blink accuracy of 97.6% and a saccade accuracy of 73.4% of the smart eyeglasses and these results were comparable to those obtained using surface electrodes. However, the authors examined the eye movements of healthy volunteers, and their findings provide a hint that the acquisition frequency of the smart eyeglasses might be adequate to capture eye movements. Furthermore, in the study of Barbara and Camilleri (2016), study participants were instructed to look at certain positions on a computer screen, while in our study, patients were not explicitly guided where to look.

The J!NS MEME smart eyeglasses showed moderate test–retest reliability to detect eye movements during PE and MI of a goal-directed upper limb motor task with overall ICC values of 0.53 (95% CI 0.29–0.76) for PE and 0.51 (95% CI 0.26–0.80) for MI (Koo and Li, 2016). Here, the explanation for the relatively low values could be the different strategies patients used for task execution. Furthermore, spontaneous eye movement patterns must have been considered as they could have overlapped task-related eye activity and therefore influenced our results. Both voluntary goal-directed (specific movement of the eye) and involuntary (blinking) eye movements were captured at the same time. Blinking frequency can vary according to daytime, fatigue, and concentration level (Wu et al., 2014; Kanematsu et al., 2019; Howell et al., 2020). In the present study, we could not ensure that both measurements were conducted at the same time of the day. In addition, our patients might have received other therapy sessions before the measurement, which could have further influenced their tiredness level and therefore their eye movements.

In accordance with the present results, previous studies have demonstrated similar reliability values. Howell et al. (2020) reported moderate test–retest reliability (r values ranged between 0.4 and 0.6) of eye-tracking assessments, and Skinner et al. (2018) presented similarly moderate reliability values (ICC −0.31 to 0.71) of eye-tracking with the Eyelink 1000 Eyetracker System (SR Research Ltd., Ottawa, Canada).

Our findings showed a large variability of ERD/ERS changes during MI and PE of a grasping task. MI- or PE-related ERD/ERS were found in individual cases; however, not in all patients MI or PE related ERD/ERS could be described. Hence, ERD/ERS changes were not comparable with earlier findings and there were no significant similarities between ERD/ERS values during MI and PE (Beisteiner et al., 1995; Pfurtscheller and Neuper, 1997; Cremades and Pease, 2007; Nam et al., 2011; Tabernig et al., 2016; Wriessnegger et al., 2018; Daeglau et al., 2020). In one electrode location (F4), the ERD/ERS values during MI and PE became more similar over time. The reason could be that once the movement sequence is learned it is performed more automatically. A possible way to handle large variability of EEG signals during MI could be the use of classification algorithms that are mostly used in studies with a brain–computer interface to improve real-time processing (Ma et al., 2022).

In the present study, task duration was not predetermined. The duration changed depending on the patients as they were not instructed how long they should maintain imagery or perform the physical task; thus, the time elapse varied. We selected a 4-s epoch that should be sufficient to differentiate different imagery tasks (Neuper et al., 2006). The defined 4-s epoch was analyzed in every patient and for each task. Average ERD/ERS values were calculated in the fixed time window. The fixed interval could have influenced ERS/ERD variability as well.

Furthermore, an automatic baseline selection was introduced for each task (1 s later than the audio cue started), compared with other methods, when taking an individual baseline during every single trial (Pfurtscheller et al., 2006). Additionally, there were no breaks between individual tasks. That could result in an overlap in the EEG signals regarding the different tasks patients were supposed to carry out such as listening and understanding the audio cue, processing visual cues, as well as the preparation for mental or physical execution of the tasks. The ERS/ERD measurements could be improved by adapting the applied paradigm and introducing longer inter-event intervals (Pfurtscheller and Lopes da Silva, 1999). In addition, the introduction of longer task sequences (controlled duration) might improve data quality. However, brief imagery can result in a larger power decrease and thus might be easier to detect (Nam et al., 2011).

Kind of movement and imagery tasks and their influence on study results could be considered. Compared with other investigations in EEG changes during MI when a simple single joint movement (e.g., finger tipping) was examined, in the present study, a complex movement of daily living was explored. The coordination of a complex task might involve more cognitive processes resulting in more EEG activation signals.

Individual differences in imagery-related ERD/ERS changes were already described in the literature (Annett, 1995; Curran and Stokes, 2003; Solodkin et al., 2004). If there is no specific instruction on how to engage in MI, patients might use different MI perspectives (internal, external) and modalities (visual, kinesthetic) requiring different brain activation patterns or strategies. In the current study, patients were free to choose their MI varieties. Here, the control of MI parameters might improve inter-subject similarities.

The ANS can provide valid and reliable measurements of MI (Collet et al., 2013). Therefore, we hypothesized that both physiological parameters HR and SpO₂ would change depending on the condition. Our linear mixed model revealed HR to be significantly different between the imagined condition compared with the resting condition and between the active and the resting conditions. Furthermore, the imagined condition showed no significant difference to the active condition and both provoked very similar responses. These findings could imply that the active and imagined conditions underlie the same cognitive pathways that are reflected in the ANS response.

However, the measured SpO₂ showed no significant difference for any of the conditions and remained constant during all three conditions. A possible explanation could be that SpO₂ remains unaffected for low-intensity demands. Even for small responses, higher loads are necessary (Daglioglu et al., 2013). Here, our motor task might not have been demanding enough, to evoke a distinct reaction.

For the present study, HR was successfully identified as a physiological readout for MI performance but not SpO₂. These results are in accordance with similar studies from Oishi et al. (2000) and Papadelis et al. (2007), who investigated ANS responses in healthy individuals. Further research should focus on MI tasks that are more demanding and hence provoke a stronger response. When the cardiovascular system is of interest, the more advanced and sensitive method of heart rate variability should be considered as a measurement parameter. It contains both the classical heart rate and its variability, which could provide a more comprehensive readout.

Prior to the present patient study, a technical validation of the setup and the experimental motor task were carried out that allowed to adjust the experiment to the needs of patients after stroke. Previous studies described detection of the eye movement in simple one-dimensional, e.g., wrist flexion and extension movements only (Heremans et al., 2008) compared with our study that included patients, who completed a complex activity of daily living tasks. Furthermore, most of the studies investigating eye movements during MI included healthy participants only. To our knowledge, the present investigation is the first study that involved patients after a stroke and evaluated their eye movements during MI compared with PE. Furthermore, our patients could select their preferred movement speed in both conditions MI and PE; thus, the spontaneous eye movements could be recorded in comparison with other studies with a strictly predetermined protocol (Heremans et al., 2012a,b).

The strength of the study is the numerous repetitions of the task in both conditions that resulted in sufficient data for the EEG analyses. Patients had to repeat each task 42 times, as it was recommended in previous investigations (Neuper et al., 2006; Pfurtscheller et al., 2006). Furthermore, an object-directed MI task (grasp and replace a cup) was used that could lead to a larger cortical activation (Li et al., 2015). Our EEG data pre-processing and analyses followed standardized procedures based on recommendations in the literature (Pedroni et al., 2019).

A limitation must be considered when interpreting the EEG findings. The MI and PE tasks might have been too short and could have led to an overlap of the activation of different brain regions. Thus, results have to be interpreted with caution.

In the present study, patients' medication was not recorded. However, the measurements took place within 7 days and it can be assumed that medication did not change in this short time period.

As MI is a mental simulation process, no external monitoring exists to control MI engagement and performance. In the present study, an attempt was made and received already promising results for the physiological parameters of HR. To our knowledge, it was the first clinical study that investigated HR to control for MI engagement in patients after a stroke.