Test on reducing incorrect muscle activations in intelligent

According to the section 2, three statistical tests were calculated to evaluate project hypotheses. First test can be considered as proving that in intelligent mode, there is a lower ratio of incorrect activation than in the non-intelligent mode. For statistical purposes, the statistical hypotheses were stated as follows:

Two populations were considered:

Both populations are not normally distributed. This leads to calculating the Mann-Whitney U test. The test statistic was U = 48,628,959.5, and the p = 1.2e-05, which rejects the H₀ hypothesis, so the populations are not equal. An additional test has been calculated to prove which population is greater in that case. Hypotheses were stated as follows:

The calculated statistic was U = 48,628,959.5, with the p = 6.11e-06, which rejects the H₀ hypothesis and implies that the share of incorrect activations during mode 1 is greater than in mode 2. The final conclusion is that the developed automation based on EMG analysis [29] statistically lowers the number of samples with EMG incorrect activation in intelligent mode. Therefore, the use of the presented intelligent mode is beneficial for a user tending to activate muscles correlated with motions supported by the exoskeletons at an adequate level, combining conventional and cognitive muscular therapy with biofeedback [38,39].

Test on increasing share of correct muscle activations in VR mode

This test was aimed at investigating the relationship between correct muscle activations during rehabilitation using the VR mode compared to the non-intelligent mode. For statistical purposes, the statistical hypotheses were stated as follows:

The hypothesis analysis was simplified to a single step compared to the previous test. However, it can also be analysed step by step as presented beforehand. Test populations were considered as:

Both populations are not normally distributed. This leads to calculating the Mann-Whitney U test. The statistics is U = 48,421,185.0, and the p = 6.25e-49, which rejects H₀ hypothesis. A boxplot of both populations is presented in Figure 4.

The final conclusion is that the project hypothesis regarding the correctness of EMG activation during VR mode was accepted. Therefore, it can be assumed that a physiotherapist controlling the motion through the VR system with constant feedback on the patients' muscular activation, intuitively performs the motions to encourage correct activation. In case of ineffective motions, the specialist can change the path to get better results. For this reason, it is worth implementing such control possibility into the systems for minimally supervised distant physiotherapy [18]. Such a functionality will be particularly beneficial after automatic emergency stops of the machines, after detecting abnormal muscular contractions, or while the patient will not activate muscles effectively enough [29].

However, it is worth considering that the VR mode also resulted in motions with slightly lower dynamics. Therefore, the muscles worked more freely and were engaged longer than in the case of a faster movement in non-intelligent mode. For this reason, the therapy motions aided with teleoperated robots can be expected to be slower than the ones recorded onsite with the physical device.

Test on increasing the share of demanding exercises in the intelligent training mode

This test was performed to confirm whether the intelligent mode increases the patient's perception that the exercise is demanding. This assessment was based on the EEG-based detection of emotional state assigned to the demanding exercises (strain state) [29]. The shares of such states detected within every exercise repetition were compared for non-intelligent mode (mode 1) and during control via VR (mode 3). For statistical purposes, the statistical hypotheses were stated as follows:

Test populations were considered as:

Firstly, H₀ was rejected with statistics U = 711,081, and p = 0.002. Therefore, an additional test with alternative hypothesis H₁ was performed. It showed that Population 1 is significantly greater than Population 2, rejecting the H₀ hypothesis with p = 0.0008. However, based on the high medians for both populations, it can be assumed that most of the repetitions were physically demanding for the patients.

In summary, the main conclusion is that both mode 1 and mode 3 are physically demanding according to the EEG emotion detection. However, mode 3 statistically decreased the share of demanding exercises. This can be a result of the patient strain states detected due to the psychological stress during the use of mode 1. This is relatively probable, as the mode 1 repetitions were always performed prior to the mode 3. Therefore, patients were not as familiar with the motions as during the further stages. Moreover, they were not used to the robot support as much as during the later repetitions. Nevertheless, patients did not report such conscious stress.

Although the hypothesis was rejected, the results of detecting demanding exercises are satisfactory and indicate the patient's correct engagement during exercises with the robot.

Motion accuracy assessment results

After the validation of three main hypotheses, the accuracy in following recorded motions was validated. The first and second tests focused on the trajectory accuracy during mode 1 (non-intelligent mode) and mode 2 (intelligent mode). Mean absolute errors were calculated as the difference in the angular position set by the main controller and the position recorded by the encoders. The trajectories were synchronized in time with one another by shifting one of them by the extreme value of the signals' convolution. Results were analysed for every rotational joint of the exoskeleton (DOF) separately. The results are presented in the Table 2 for test 1, and in the Table 3 for test 2.

The results demonstrate very high accuracy in motion following, especially for robots interacting physically with humans [40]. The highest average inaccuracies occur in DOF 1 (hip abduction/adduction combined with external/internal rotation) and DOF 3 (knee flexion/extension). Inaccuracies in the former joint are the result of a different, more comfortable patient position during independent repetitions with the robot, compared to the reference motion recording. Moreover, it was observed that the latter joint has higher inaccuracies than the others due to the characteristics of the exercises performed, especially if the foot was moving very close to the floor. Potential friction of the foot sole during motions brought additional inaccuracies into DOF 3 and 4.

The second set of two tests investigated position accuracy in mode 3 (VR mode). Test 3 examined the accuracy of the trajectory set by the VR controllers and its mapping onto the real trajectory. The calculation technique corresponded to the one used for the first two tests. The aim of this was to verify how precise the control interference based on the digital twin in VR can be. The results are presented in Table 4 while the data boxplot is visualised in Figure 5a. In general, the quality of control was visibly lower than in the non-intelligent mode. However, the errors typically did not exceed 5 degrees, where their median remained in the range of 1-2 degrees. Such results are fully acceptable for the analysed case of telerehabilitation, especially considering the potential accuracy of alternative control signals [18,41].

It was observed that the physiotherapist operating in virtual reality gained experience in operating the robot with each session. The highest average inaccuracies were recorded for the ankle flexion/extension (DOF 4) - responsible for orientation for the most distal segment of the exoskeleton's kinematic chain. This could result from the backdrivability of the corresponding drive and the least interest in controlling this segment, as it is often preferable to remain comfortable for the patient rather than fully controlled [42].

Test 4 aimed to verify how accurate the rehabilitation performed via the VR-based digital twin is compared to the reference trajectory. Therefore, it was a check of how accurately a physiotherapist can execute fully-planned movements with a robot. For this experiment, the reference trajectory was overlayed on a digital twin in VR as a semi-transparent phantom of the extremity. The physiotherapist was trying to follow the phantom with the digital twin of the exoskeleton and the extremity coupled with their physical representations in the control loop. The obtained real-life motion trajectories were compared with the reference trajectories to compute inaccuracies, as in all three previous tests. Results of this analysis are presented in Table 5 and the boxplot in Figure 5b.

The inaccuracies from test 4 are the highest compared to the previous cases. The most distal and proximal degrees of freedom (1 and 4), less involved in the test motions, have the best results, while the ones engaged more - knee and hip flexion/extension - have errors exceeding 10 degrees.

Nevertheless, the errors were decreasing while operating the system. Therefore, it can be expected to minimize them by educating physiotherapists and medics in the field of using a VR environment, becoming more common [43]. Moreover, test 3 presented that the system is capable of performing control with the fully acceptable accuracy, which should be the target of the mentioned minimization.

To contextualise the reported errors, the recorded motion trajectories span within the range of 30–90° per joint depending on the exercise type and the individual participant’s range of motion for the primary therapeutic joints (DOF2: hip flexion/extension; DOF3: knee flexion/extension). It is assumed that the whole set of experimental exercises covers the full range of mobility for a participant. The precise values for every joint for a healthy participant are presented in Table 1.

Resultant mean absolute errors of approximately 0.67° and 1.9°, respectively, in Mode 1 correspond to relative errors below 6% of full motion amplitude, which is acceptable for the application in human-interactive rehabilitation robots. The elevated errors in DOF1 are primarily attributable to the difference in resting patient posture between the reference-recording session and the therapy session, rather than to control or actuator limitations. Errors in DOF3 and DOF4 are partly attributable to foot–floor friction during exercises in which the foot approaches the ground surface.

Assessment of safety functionality for extreme cases

Separate research on emergency systems was performed, both for EMG and EEG signals, triggering the stop. In order to provide maximum safety for the patient using the exoskeleton, it was decided to implement the entire service of EMG-based muscle contraction detection and EEG-based pain recognition in the local electrical components at the patient’s side (Raspberry Pi microcomputers - RPi) of the robotic system. In case of performing remote rehabilitation, a VR operator sees the corresponding messages of detected emotional states and muscle contractions. Additionally, a critical message appears upon cramp or pain detection. When any of the conditions indicating patient hazards (cramp or pain) are detected, the robot's emergency procedure is immediately initiated.

Data from the EEG headset and EMG sensors are recorded and stored by the database server hosted on the RPi located in the electrical cabinet of the exoskeleton. If any of these signals indicate the prediction of soon-to-be pain or cramp [29], an emergency interruption flag is immediately sent to the main control computer to stop exercise. All devices are located in the patient's location, ensuring minimal delays, which is 1 ms via Ethernet cables. For the safety system trials, both cramp and pain states are triggered randomly during trials with the electrical signals emulating real-life muscular or brain activation.

Firstly, safety test was conducted for simulated spastic cramps. For this, ten sessions of cramp detection have been performed. They were detected by analysing the first derivative of EMG signals registered for every muscle separately. The cramp was assumed as potentially possible while the individual threshold was exceeded for the defined time. Every measured delay between detecting a cramp by the RPi controller and the robot stop has been presented in the Table 6. It was assumed that the additional delay of measuring and processing the EMG was not bigger than 100 ms in total (based on the setup frequency) [29].

The cramp detection formula is based on the backward differentiation formula (also known as the backward Euler method). If a relatively high EMG value is registered, compared to the previous one, and if the value of the derivative is higher than a threshold, the safety procedure is activated. Delays are calculated as a D-C column timestamp subtraction (see Table 6). All the registered delays were lower than 25 ms. This value was relatively low compared to the time of computation and measuring the signals. That was possible to achieve only by handling the security procedure locally, which maximizes the safety and robustness of the system.

Then, ten sessions of pain detection have been performed. The pain states detected based on bandwidths from selected frontal lobe electrodes were reflecting the states in which there is a high risk of real pain appearing within two seconds [29]. Every measured delay between receiving information about a pain state by the RPi controller and activating the emergency state has been presented in the Table 7. It was assumed that the additional delay of measuring and processing the EEG was not bigger than 2 s in total (based on the setup frequency) [29].

All the registered delays were lower than 45 ms. Like for cramp detection, this value was relatively low compared to the time of computation and measuring the signals. The reaction to risks detected by the EMG and EEG signal was very fast and met the design assumptions for the SmartEx-Twin exoskeleton. Therefore, they are applied to the final design of the system.

Study limitations

Several limitations should be considered. First, the cohort was intentionally heterogeneous, spanning healthy volunteers and participants with neurological, orthopaedic, cardiological and undiagnosed conditions across a wide age range. This composition reproduces a realistic rehabilitation setting, but the subgroups were too small to support a stratified analysis. The pooled comparisons, therefore, cannot attribute effects to specific conditions, and the results should be read as a feasibility-level characterisation across a mixed population rather than as condition-specific evidence.

Moreover, all teleoperation was performed by a single physiotherapist whose accuracy did not improve at the statistical level over the course of the study. Nevertheless, it is possible that various operators can perform with changing performance, what should be considered in the planned longitudinal continuation of the investigation.