1. Introduction
Rehabilitation robots, or exoskeleton robots, detect user intentions and provide assistance by attaching robotic limbs to humans. Along with the development of robotics technology, there has been active research into the design, structure, and control of rehabilitation robots [
1]. Research on rehabilitation robots is advancing rapidly in the rehabilitation, medical, and power augmentation fields [
2,
3]. Concurrently, the market demand for rehabilitation robots is growing [
3]. However, as this field develops, issues concerning the comfort of human–machine interfaces are becoming increasingly significant [
2]. At present, there is a lack of objective evaluation criteria and guidelines for assessing the comfort of rehabilitation robots. In rehabilitation robotics, applying force to control the robot and maintain its attachment to the user is inevitable. However, excessive force can cause discomfort, pain, and injury from the device, and in severe cases, it can lead to skin problems such as pressure ulcers [
4]. Consequently, assessing the fit of rehabilitation robots is crucial to ensure user comfort and prevent injuries caused by excessive pressure.
The VAS has been widely used for evaluating pain or discomfort [
4,
5]. Daly et al. [
4] used a VAS as a pain score to determine a correlation with the maximum pressure of sockets of upper-limb prostheses. In their study, the VAS was normalized by the subject’s average value to minimize the threshold bias of individuals’ perception of discomfort. Meyer et al. [
5] created questionnaires that included the a VAS, the Likert scale, and a numerical rating scale to assess the usability of wearable robots and reported that the VAS, Likert scale, and open text were mainly preferred by subjects.
Several studies quantified discomfort based on multiple bio-signals, as the autonomic nervous system (ANS) is the main body in response to stimuli. However, only a few studies used EDA signals and StO
2, such as Léger et al. [
6]. In their study, they evaluated workstations, but there are only a handful of studies that applied these signals to other types of binding parts in rehabilitation robots.
Many studies show that EDA signals are the most reliable for measuring discomfort [
7,
8,
9,
10]. EDA, which is also known as GSR (galvanic skin response), refers to the electrical phenomenon caused by sweat secretion from the skin and appendages [
11]. It is measured on the skin surface, primarily on the palms, fingers, and soles, where sweat glands are distributed [
12,
13]. When the body is subjected to physical stimuli as well as various mental stimuli, such as tension and excitement, sweat secretion from sweat glands increases in response to the somatosensory and sympathetic nervous systems (SNS) [
13], which reduces skin resistance and increases the strength of the EDA signal. The EDA signal can be decomposed into phasic and tonic components. The phasic component is a rapid response to external stimuli, while the tonic component represents a slowly changing baseline of the signal [
13]. Several studies were conducted based on these characteristics of EDA. Kong et al. [
14] applied electrical stimulation to the right forearm, decomposed the EDA signal using a high-pass filter with a cutoff frequency of 0.05 Hz, and developed an algorithm to detect the phasic component and determine the pain based on stimulation intensity. Kim et al. [
15] measured EDA on the left palm by applying pressure stimulus to the left scapula, observing a linear increase in the maximum SCR (skin conductance response) amplitude as the pressure intensity increased. Posada-Quintero et al. [
16] found that the mean SCL (skin conductance level) and the number of SCR peaks significantly increased when applying physical stimuli, postural stimulation, and the cold pressor test. Hosseini et al. [
17] demonstrated that 87 features extracted from EDA signals in the WESAD (wearable stress and affect detection) dataset were sufficient to classify stress and non-stress groups with over 80% accuracy. They also found that five features—the mean SCL, maximum SCL, number of SCR peaks, maximum SCR amplitude, and standard deviation of SCR rise time—were adequate to classify the two groups with 97% accuracy.
Numerous studies explored the effects of tissue compression on blood circulation and microvasculature [
18,
19,
20]. Linnenberg et al. [
21] reported that high-pressure tissue compression during the use of rehabilitation robots can cause discomfort, pain, or soft tissue damage owing to ischemia. Their findings revealed a decrease in tissue oxygenation of 4.1% per minute at the robot’s binding parts during rest. In some research focusing on wearable robotic bindings for the lower extremities, it was observed that the thigh caused more pain and exhibited a significant increase in oxygen saturation at the same pressure compared with the shank, particularly during standing or walking activities [
22,
23,
24]. In addition, Nam et al. [
24] conducted a study to measure StO
2 and quantified the comfort of a standing supportive rehabilitation robot at different binding sites. This study employed near-infrared spectroscopy (NIRS), which utilizes near-infrared light at wavelengths of 630–1300 nm. NIRS penetrates tissues to depths of 1–3 cm and measures changes in the concentration of oxygenated and deoxygenated hemoglobin; therefore, it is effective for the continuous and non-invasive monitoring of the relevant area.
In this foundational research, we employed a pneumatic cuff for pressure stimulation—a commonly used binding component in rehabilitation robots—to analyze the effects of binding forces through bio-signals such as EDA and StO2. This study aims to elucidate the relationship between pressure and bio-signals in rehabilitation robots, potentially making a significant contribution to the design and user interface of such devices.
4. Discussion
In this study, we aimed to quantify the comfort of the binding parts of rehabilitation robots in a sitting position by using bio-signals such as EDA and StO
2 under different pressure conditions. This study employed a circular pneumatic cuff to apply continuous pressure to the binding parts, unlike previous studies that used digital algometers [
32]. Three different pressure levels were applied to the thigh to measure both EDA and StO
2. For EDA, the mean SCL, maximum SCR amplitude, and SCR counts were considered and correlated with the normalized VAS. The EDA characteristics tended to increase with increasing pressure, which is in agreement with previous studies [
15,
16], and this increase was found to be positively correlated with the normalized VAS. We found that, among the EDA characteristics, the mean SCL showed a relatively strong correlation and was statistically significant under the measured pressure conditions. Consistent with previous work, the decrease in StO
2 obtained from the NIRS sensor was positively correlated with the normalized VAS [
24]. While statistical significance was observed across all pressure conditions, the difference in values between 20 kPa and 30 kPa was less distinct than with other pressure changes. They suggested that beyond a threshold pressure of 20 kPa, the oxygen saturation in the tissue did not significantly decrease owing to vascular occlusion [
24,
33].
Some studies utilized machine learning to classify discomfort from EDA signals [
8,
9,
17]. Based on these approaches, future research should focus on using machine learning to evaluate discomfort at binding parts. Beyond EDA, various bio-signals such as EMG (electromyography), HRV (heart rate variability), and RESP (respiration) were also applied in the study [
34,
35,
36,
37]. Even though it is quite challenging to quantify discomfort at the binding parts using HRV, HRV characteristic values exhibit a clear increasing trend with pressure increase. This observation aligns with a study that successfully performed binary classification of discomfort using HRV characteristics [
38] and is further supported by research suggesting the potential of high-frequency HRV as a physiological marker in pain assessment [
39]. Therefore, it would be important to apply more in-depth analysis methods to HRV in future studies.
In this study, changes in EDA were compared by applying different pressures for either 1 or 5 min to the binding parts of participants in a sitting posture. Comparing
Table 2 with
Table 5 shows that the F-values in Experiment 2 were smaller than those in Experiment 1. This reduction may reflect the homogenization of physiological responses over a prolonged stimulus period. It can be concluded that the initial variability in EDA responses, which could be amplified owing to individual differences in physiological and psychological conditions, tends to diminish over time because of adaptation to pressure. In addition, the number of subjects in Experiment 1 (n = 13) is larger than that in Experiment 2 (n = 10), which may affect the statistical results.
In practice, rehabilitation robots are likely to be worn for longer durations compared with the experimental conditions, and it is essential to consider the wearability factor when applying actual pressure in real-world settings. Given that a longer duration of pressure application may enhance the correlation of mean SCL, this parameter could potentially serve as a valuable indicator for assessing prolonged discomfort in users [
37]. Furthermore, while our experiment involved stimulation of the right thigh only, it is important to acknowledge that actual wearable structures typically involve binding on both thighs. Since the binding parts in the actual rehabilitation robot are more complicated than the pressure cuff, it would be necessary to determine the binding parts under various conditions, such as in a standing state. Although this study did not use the actual binding parts of a rehabilitation robot, it represents a fundamental approach in which a pneumatic cuff-type binding is applied as a preliminary model.
The max amplitude of SCR and SCR counts did not show statistical significance. This may have resulted from the high dependence of the mean SCL on SCR since a high-pass filter was used in the decomposition of the EDA signal. We believe that the use of advanced EDA analysis tools, such as the cvxEDA algorithm, can significantly minimize this problem. The cvxEDA algorithm demonstrated a strong ability to detect SCRs from raw signals reliably [
40], thereby enhancing the statistical significance of the maximum SCR amplitude and SCR counts. In addition, it is important to consider that not only noise or motion artifacts but also emotional stimuli can introduce changes in the EDA signal. In this study involving 23 healthy subjects, it is acknowledged that patients with lower-limb disabilities or disorders may present different bio-signal patterns [
25,
41]. Therefore, further research including these patients—who represent the actual users of rehabilitation robots—is essential for comprehensive analysis. Such efforts will not only validate and enrich our results but also contribute to the quantification of comfort in the binding parts of rehabilitation robots.