A Data-Driven Fuzzy Logic Method for Psychophysiological Assessment: An Application to Exoskeleton-Assisted Walking

Multimodal physiological monitoring and related estimation of the PsychoPhysiological (PP) state play an essential role in investigating the physical and cognitive workload of people executing a motor task. The aim of this work was to develop a data-driven Fuzzy Logic method to estimate four PP indicators, i.e., Energy Expenditure, Fatigue, Attention, and Stress, and test it in a study including ten healthy participants walking while assisted by a lower limb treadmill-based exoskeleton. PP indicators were compared with participants’ self-reported evaluation of the human-robot interaction experience following the administration of a dedicated questionnaire. Results from a correlation analysis demonstrated that the output of the Fuzzy Logic method was consistent with the participants’ subjective assessment.

facing physical effort and focusing attention on the ongoing activity.Consequently, many processes and changes happen in their body related to both physical and cognitive spheres [2].The PsychoPhysiological (PP) state is an indicator that can account for these changes.It can be defined as the combination of the physiological state and the affective state of the person: the former is linked to the set of vital parameters, and the latter relates to the psychological aspects [3].
Several solutions have been used to monitor multimodal parameters with the aim to assess and even improve the usertechnology interaction [4].Anyhow, besides the monitoring of the raw physiological measures, the estimation of the PP state is useful to thoroughly investigate physical and cognitive states.
The assessment of the PP state can rely on physiological data such as the electrocardiogram (ECG), the respiration, and the electrodermal activity, also called Galvanic Skin Response (GSR), measured by wearable sensing units [5], [6].It was demonstrated that the Heart Rate (HR) decreases after the administration of visual, auditory and/or haptic stimuli [7].On the other hand, an increase in HR or in Heart Rate Variability (HRV), with respect to the baseline, is associated with an excited condition [8].Also the breathing activity reflects the PP state of a user even if the modifications in the Respiratory Rate (RR) are slower than the ones shown by the other physiological signals.For example, during highly intensive physical workload, RR decreases [9].Lastly, the GSR is one of the most analyzed physiological parameters to estimate the user's arousal level [10].Indeed, the electrical properties of skin change every time a person is stimulated by visual, acoustic, haptic, or physical stimuli.The higher the GSR, the higher the user's arousal and cognitive workload [11].
Typically, in human-technology interaction tasks, users are asked to rate their perceived experience by means of questionnaires [12], [13].On the other hand, the objective PP state estimation has a higher impact since it can be useful to take decisions in real-time and adapt the behavior of a device or a software application with which the user is interacting, especially in robotics applications.
In [14], PP measurements were used, in combination with performance and patient's biomechanical data, in a discriminant analysis model to predict whether the difficulty level of a rehabilitation task had to be changed.Some methods c 2024 The Authors.This work is licensed under a Creative Commons Attribution 4.0 License.
For more information, see https://creativecommons.org/licenses/by/4.0/Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
proposed in the literature include inference systems taking as input physiological measurements to modify the difficulty level of a serious game provided to patients during rehabilitation [15], [16].Another solution was developed in [17], [18] to dynamically adapt the behavior of an upper-limb rehabilitation robot according to the patient PP state, defined with arousal and valence dimensions, by means of a two-step Fuzzy Logic controller.An attempt to estimate the cognitive load of participants undergoing a treadmill-based exoskeleton rehabilitation session was provided in [19], where a linear discriminant analysis predicted the cognitive load starting from physiological measures and task performance.Literature hence highlighted that physiological monitoring systems play an essential role in estimating individuals' PP state and that this estimation is suitable to enable the development of robotic systems adapting according to users' real-time needs.
In this work we were interested in four relevant PP indicators to assess physical and cognitive activities: Energy Expenditure (E), Fatigue (F), Attention (A), and Stress (S).The E assesses the physical workload required from the user to accomplish an assigned task [20].Typically, it is quantified by measuring the O 2 consumption based on indirect calorimetry.Machine learning algorithms, such as decision trees, were already used to predict the E of participants starting from physiological and demographic information [21].
Users undergoing prolonged physical and/or mental engagement may experience F conditions.Few studies in the literature addressed the F estimation starting from the physiological measures.In particular, the F level of participants undergoing cycloergometry sessions was computed with a support vector machine approach in [22].
A and S levels can be used to assess the cognitive workload with positive and negative acceptations, respectively.The A level is assessed in literature by integrating physiological data with electroencephalography or pupillary response.A dynamic Bayesian network was used in [23] to distinguish between different A levels of the participants.S is the most investigated PP indicator.Many studies have been conducted in recent years to analyze stress conditions in healthy subjects.Most of them aim at designing machine learning classifiers, like logistic regressions models, decision trees, and support vector machines, capable of distinguishing "stressed" from "notstressed" condition [24], [25], [26].
Supervised classifiers are the most commonly used approaches to predict PP indicators due to their high estimation accuracy.Anyhow, they need to be properly trained with well-defined and structured data to be acquired and labeled by eliciting desired PP states.This process can be hardly pursued since evoking extreme values for the PP indicators of interest is particularly complex and not always possible, and also labeling can potentially have limited reliability.Lastly, classifiers are only capable of providing discrete estimations, or even merely binary states [27].Fuzzy Logic is an effective approach to overcome the highlighted limitations.Indeed, it does not rely on the labeling of predefined states purposely elicited in dedicated training tests, and also it provides continuous estimates of the variables of interest.Studies exploiting Fuzzy Logic methods were only focused on estimating users' arousal and valence based on physiological measurements to model affective states during interaction with entertainment technologies [28], [29].Traditionally, Membership Functions (MFs) and rules linking input-output data, that are needed for the design of Fuzzy Logic models, have to be manually defined by the experimenter based on personal experience and insights about the phenomena under investigation.This dependence on the experimenter's choices is a limiting factor, compromising the thoroughness of results, and can be overcome if an automatic and data-driven methodology is established.This paper aims at proposing a new method based on Fuzzy Logic approach for estimating PP indicators during physical tasks based on physiological data.The development of a unified model capable of providing different information about the users' PP state allows a multimodal assessment since both physical and cognitive workloads can be estimated and interpreted during the execution of physical tasks.More specifically, relevant contributions of the work include: ii) the automatic data-driven definition of the MFs and the development of a set of Fuzzy rules based on literature knowledge of relationships among physiological information and PP indicators; iii) the continuous estimation of PP indicators to potentially provide useful time-varying feedback on the user' state; iv) the demonstration of the method in a specific application entailing exoskeleton-assisted treadmill walking.No previous similar examples of this application are available in the literature to the best of the Authors' knowledge , hence we aimed to fill this gap.Indeed, the lack of recognized assessment methods and available structured datasets for this scenario made it a worthy testbed for our research.Moreover, even if exoskeleton-assisted walking with body weight support and on a treadmill is not inherently quite complex, especially if no specific challenging feedback-based objective is set, understanding the PP aspects of users can be anyway crucial.Indeed, in applications such as robot-assisted rehabilitation, where potential monotony and inattention can impact patient's engagement, motivation and, in turn, therapy outcomes, continuous assessment of the user experience is significantly valuable.
Continuous predictions offer a multitude of advantages when compared to discrete estimations of PP processes.They provide a real-time view into a patient's physiological and psychological states, allowing for a more comprehensive and detailed understanding of their responses to rehabilitation exercises.This ongoing insight empowers therapists and robotic systems to make well-informed, timely decisions in adapting rehabilitation strategies to comply with each individual's unique needs.Furthermore, real-time PP estimation holds promise for various applications, including integration into visual feedback systems [30], [31], and development of control strategies to further personalize motor therapy by suiting the precise requirements of individual users [32], [33], [34].Nonetheless, the method has general validity and has the potential to be applied to several scenarios in which participants undergo physical and/or cognitive workloads.Indeed, the data-driven nature of the approach allows new data to be recorded under different experimental conditions to obtain a Fuzzy Logic model tailored to specific applications.The paper is organized as follows: Section II reports the materials and methods needed to implement the proposed approach and describes the performed testing experiments.Section III presents and discusses the obtained results while conclusions and future work are drawn in Section IV.

II. MATERIALS AND METHODS
In Figure 1, the functional scheme of the proposed approach for estimating the PP indicators is shown.The Fuzzy Logic method is capable of mapping highly variable inputs, such as the physiological ones, into the four PP indicators of interest, namely E, F, A, S. The method has general validity and can be exploited to estimate the PP state of users interacting with any robot and/or executing any physical task.To test our approach, an experiment is designed including five healthy volunteers equipped with a physiological monitoring system and walking with the aid of a lower limb treadmill-based exoskeleton.
In the following sub-sections, all the elements introduced in the block scheme of Figure 1 are described in detail.

A. Psychophysiological Indicators
In this paper, the four aforementioned PP indicators (E, F, A, S) are used to assess user experience during exoskeleton-assisted walking.These indicators are briefly recalled in this Section to highlight their relationship with physiological measures and to contextualize their use in the selected application scenario.
E plays a pivotal role in all the tasks involving a human being in physical activities [35].The E assesses the metabolic needs of a user and is measured in general by means of calorimetric techniques [36].Anyhow, it can be also indirectly estimated by using other physiological measures collected with wearable sensors.The parameters showing the widest modifications during high E conditions are HR, HRV, and RR [37].In the exoskeleton-assisted walking context, E is related to the physical workload required by the user to walk as well as to the perceived comfort.Indeed, an uncomfortable robot may hinder the user's movements thus causing a physical activity more intense than needed and resulting in a high value of E.
F indicates the disinclination to continue an ongoing task whether its nature is physical or cognitive.Hence, it increases with the overall task workload.The increase of F has an impact on HRV, in particular on the absolute power of the lowfrequencies [38], as well as on the RR [9].Since F takes into account both the physical and cognitive workload, during a human-robot interaction session it can be used as an indicator of the user's tendency to keep on adopting the technology (e.g., a patient undergoing robotic rehabilitation may be prone to abandon it in case of high F values).
A represents the cognitive load experienced in performing a certain task.Several studies estimate A starting from non-invasive physiological measurements by challenging the enrolled participants with tasks of increasing difficulty [39].The user's physiological responses revealed some identifiable patterns in heart electrical activity [40], respiration and GSR [41].The A level can be used, during exoskeletonassisted walking, as an indicator to assess the user's cognitive workload, with positive meaning, and involvement.High values of A indicate that the robot use prompts the user to be focused on the activity to be accomplished.
Lastly, S is defined as a cognitive workload with negative acceptation.The most informative signal for its detection is the GSR [42].However, modification of the overall physiological state can be observed in stressed conditions due to the intense sympathetic activities influencing both the cardiac and respiratory functions.During an exoskeleton-assisted walking session, S represents a measure of the perceived frustration and it reaches high values when the robot use is demanding.

B. Physiological Monitoring System
1) Physiological Sensors: Wearable sensors are used within this paper to measure the ECG, the respiration activity, and the GSR.The heart and respiratory activities are monitored by using the BioHarness 3.0 chest belt, developed by Zephyr TM Technology.This device fuses i) capacitive sensors, detecting the potential variations on the chest of the users due to the electrical heart activity, to measure ECG (sampling frequency: 250 Hz) and i) stretch sensors, monitoring the deformations imposed by the expansion and compression of the rib cage, to measure breathing activity (sampling frequency: 25 Hz).The GSR is measured by using two electrodes of the Shimmer 3 GSR+ Unit, placed on the index and middle fingers of the non-dominant hand.Such a difference in potential allows for retrieving information about the user's electrodermal activity (sampling frequency: 52.1 Hz).Both the BioHarness 3.0 and the Shimmer 3 GSR+ Unit provide wireless communication with a development computer.A custom graphical user interface allows data synchronization, recording, and storage.Moreover, some physiological parameters, described in Section II-B2, are computed in realtime and stored in a text file at a sampling frequency of 25 Hz.
2) Physiological Parameters: Given the raw physiological data measured by means of the aforementioned wearable sensors, several parameters can be extracted.
One of the most important aspects of ECG signal processing is the identification of the R wave, which is easy to be detected thanks to its structural form and high amplitude.If the R-peaks are detected from the ECG signal, the interbeat interval (IBI H ) can be computed.Moreover, based on the R-peak it is possible to identify the whole QRS complex.The IBI H , also called tachogram, represents a temporal series of RR peak intervals, i.e., the time elapsed between two consecutive R-peaks.Starting from the IBI H signal, it is possible to compute the instantaneous HR, expressed in beats per min (bpm), as: where i indicates the i-th detected R-peak.
In addition, both time-domain and frequency-domain HRV metrics are computed from the IBI H signal.The Root Mean Square of Successive heart beat Differences (RMSSD), expressed in ms, is a time-domain parameter computed as: where N = 2 f s is the number of samples in a time window of 2 s, f s is the sampling frequency, i and j indicate the i-th and the j-th sample respectively.
The ECG analysis in the frequency domain requires the Fast Fourier Transform (FFT) of the IBI H signal to extract the absolute power (in ms 2 /Hz) of the low-frequency band f ∈ (0.04 − 0.15) Hz.A recording of at least 3 min is required to compute the FFT of the IBI H signal [43].The Low Frequency (LF) power can be computed from the FFT of 3 min IBI H recording as: where N = 180 f s is the number of samples in a time window of 3 min.From the respiration signal, the respiratory events are detected by identifying the local maxima, which represent the largest expansion of the rib cage.As for the IBI H computation, an inter-breath interval signal (IBI R ) can be defined.The RR, expressed in breath per minute (bpm), is computed as: where i indicates the i-th collected sample.
The GSR signal can be used to extract the two main components: the tonic level and the phasic response.By applying a 4-th order zero-lag Butterworth low-pass filter, with a cutoff frequency of 0.1 Hz, it is possible to compute the Skin Conductance Level (SCL).It represents the tonic level in the absence of any particular environmental event.The other component of the GSR, i.e., the Skin Conductance Response (SCR) is an event-dependent, phasic, and highly responsive parameter and can be found in the 0.1 − 5 Hz frequency band [44].Both the SCL and SCR are expressed in μS.
Overall, the i-th physiological vector extracted during a recording can be defined as: 3) Data Normalization and Filtering: All the collected data require a normalization procedure to allow comparisons among different participants.Indeed, physiological measures exhibit high intra-and inter-subject variability as a result of age, gender, time of day, and many other factors.Normalization can reduce the effect of such variability by evaluating the response X R of each physiological parameter with respect to a baseline condition (Resting Baseline, RB), i.e., considering the participants sit comfortably, blindfolded and acoustically isolated for 5 min.The normalization procedure is pursued as follows: where i indicates the i-th sample, X represents the current physiological vector and X RB is the mean value of each physiological parameter collected during the baseline phase.
A moving average filter with a window size of 2 s is then applied to the computed X R to smooth the collected data without loss of relevant information.Indeed, our parameters of interest are slowly changing so that the moving average filter is employed only to remove undesired artifacts or noise [45].The processed physiological parameters are available in realtime at a frequency of 25 Hz, as previously explained in Section II-B1.

C. Fuzzy Logic Method
The physiological responses X R are used as input signals to the Fuzzy Logic method.It needs the definition of its input, outputs, MFs, and rules.The MFs transform the membership of a specific element into a percentage membership in the set.They weigh each input signal included in X R , defining overlaps between the input levels and determining the PP indicators.For each input signal of X R , the MFs are generated by using the data acquired from all the enrolled participants.The collected data are grouped in such a way as to highlight three levels of activation for each physiological parameter defining the linguistic variables "LOW", "MID", and "HIGH".The "MID" value is built as a triangular function as: In (7), a and c locate the feet of the triangle, b finds the peak and x represents the input data.On the other hand, "LOW" and "HIGH" functions are trapezoidal, to remove fuzziness from the extreme input values, and are expressed as: Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.The approach proposed in this paper requires the definition of the MFs keypoints, i.e., {a, b, c} for the triangular function and {a, b, c, d} for the trapezoidal one, in a datadriven manner.This means that the MFs are defined based on the data acquired during the experimental sessions of exoskeleton-assisted walking.In particular, all the collected physiological responses X R are sorted in ascending order and divided into three subgroups, each containing the same number of occurrences.For the sake of brevity, the response of the physiological vector X j,n R is introduced, where j indicates the parameter j ∈ {HR, RMSSD, LF, RR, SCL, SCR} and n represents the MF (n = {L, M, H}, for "LOW", "MID" and "HIGH", respectively).For each physiological response component, the keypoints of the three MFs can be computed by using statistical features of the different activation groups, i.e., the mean value and the standard deviation, as reported in Table I.In our notation, the min(•) and max(•) functions are used to find the smallest and the biggest elements in a dataset respectively, and mean(•) computes the mean value of the dataset.Similarly, the Fuzzy Logic method requires the definition of the MFs of the output variables, i.e., the PP indicators.Each PP indicator is identified by using three triangular MFs meaning "LOW", "MID", and "HIGH" activation.The keypoints used to implement the output MFs are listed in Table II.The equally spaced placement of the membership function centers ensures a uniform partitioning of the full range of values.It allows the model to represent different PP states in a gradual manner, such as "LOW", "MID", and "HIGH", reflecting a smooth progression.A similar approach has also been pursued in [17].
Once the MFs are created and the fuzzification module is completed, the fuzzy rules can be defined and implemented.Sets of 11, 9, 17, and 17 rules are defined for E, F, A, and S respectively, for a total of 54 rules.Such rules are extracted from a careful analysis of the scientific literature [11], [37], [46].The entire set of conditional rules implemented to transform the physiological responses X R into the PP indicators is reported in the Appendix A.
Lastly, to finalize the design of the Fuzzy Logic method, the fuzzy operators have to be defined.The Mamdani technique has been selected.The logical AND and implication methods are implemented with the min(•) function.The logical OR operator and aggregation methods are defined by the max(•) function.The implemented defuzzification process is the area centroid or center of gravity.The MATLAB Fuzzy Logic Toolbox is used to build up the method.

D. Exoskeleton-Assisted Walking
To prove the effectiveness of the proposed method in estimating the four PP indicators, a study on healthy participants has been carried out.In the following, the experimental setup and protocol are presented.
1) Experimental Setup: A representative participant wearing the lower limb exoskeleton and the physiological monitoring system is shown in Figure 2. The exoskeleton used in the experiments is the Lokomat R Pro by Hocoma [47].It is a bilateral robotic orthosis that is used in conjunction with a treadmill and a dynamic Body Weight Support (BWS) to assist the patient's lower limb movements in the sagittal plane.The BWS is employed to reduce gravitational forces, ensure safety, and maintain the user balance.The Lokomat R hip and knee flexion/extension joints are assisted by linear back-drivable actuators.Passive foot lifters support ankle dorsiflexion during the swing phase.The limbs of the user, which are fixed to the exoskeleton by straps, are guided by means of a compliance controller, whose reference trajectory is defined by the physical therapist together with the walking speed.The level of assistance is decided based on the needs of the user and is increased by selecting a higher level of stiffness in the controller.The higher is the stiffness, the higher is the level of assistance and the less is the user's contribution to the task.The assistance level can be regulated from 0% (theoretically entailing no robot contribution) to 100% (providing full assistance, similarly to what is pursued with a position controller).
The wearable monitoring system, presented in Section II-B, is used in the experiments to capture many physiological signals of the user that are not always correlated with each other.It is therefore possible with a reduced set of sensors to get a precise picture of the user's physiological state.The data analysis used to compute the physiological parameters has been described in depth in Section II-B2.
2) Experimental Protocol: Ten healthy participants (29.9± 10.6 y.o., 4 males and 6 females) are enrolled and have given their written consent to participate in the study.The experiments are performed at the Neurorehabilitation 1 Department, Fondazione Santa Lucia (FSL), Rome, Italy.The experimental Each enrolled participant had no previous experience with the Lokomat exoskeleton.This ensures that no habituation effects could alter the results of our experiments.Each participant is equipped with the wearable physiological monitoring system described in Section II-B1, i.e., the BioHarness chest belt, and the Shimmer 3 GSR+ Unit.When the experiment begins, a 5 min RB is recorded.In particular, the participants are asked to sit comfortably, blindfolded, and acoustically isolated to ease her/his rest condition.The baseline recording is performed in the sitting condition of the participants since changing body posture from sitting to standing wearing the Lokomat was demonstrated not to affect the physiological state [48].The mean value of the physiological features extracted in these 5 min is used in (6) to compute the physiological responses X R .
According to the study protocol, after the RB the participant wear the Lokomat R and starts the walking session.Before the recording phase, a warm-up period of at least 15 min is performed.During this phase, the physical therapist set the desired joint kinematic patterns for the lower limbs, as well as the gait speed on the treadmill, the level of assistance, and the BWS to be delivered.The range of possible values for Lokomat R is defined a priori based on the experience collected in daily clinical practice.In particular, the gait speed is selected in the range 1.3 − 1.5 km/h, the assistance level is in the range 70 − 90 % and the BWS is in the range 30 − 50 %.In a recent study we have demonstrated that changing, even largely, the robotic assistance from the Lokomat does not significantly alter walking performance in terms of biomechanical parameters [49].Hence, it is worth highlighting that the selection of the assistance level had not an effect on results.Moreover, the selection of gait speed and of BWS values was pursued in a range narrow enough to avoid introducing differences in the walking performance among participants (as demonstrated again in [49]).During the warm-up phase, the physical therapist iteratively adjusted the parameters based on constant verbal feedback from the participant, to identify the most comfortable walking conditions.When the most suitable settings are reached, they are kept constant and a 15-min phase measurement starts.The physiological responses extracted in this walking phase are used to compute the four PP indicators of interest.At the end of the measurement session, a custom questionnaire, detailed in Section II-E, is administrated to the participants.

E. Subjective Assessment
The Authors have recently developed a novel multifactor questionnaire for assessing the user experience in interacting with lower limb exoskeletons [13].It includes a variable number of items to be rated on a 7-point Likert-type scale (from 1"I strongly disagree" to 7 "I strongly agree").In the experiments carried out in this paper, 8 items are selected to assess the participants' self-reported evaluation of the robot use.In particular, two items (I PP1 and I PP2 ) are chosen for each PP indicator under analysis: • I E1 : Compared to the results I can achieve, using the exoskeleton is too physically demanding.The item assesses how much hardly the participant had to work to achieve the required results; this item also referred to the specific task performed with the exoskeleton.• I E2 : I had to work hard to reach my level of performance.
The item assessed the degree to which the participant believed that using the exoskeleton helped him to achieve suitable performance.• I F1 : Using the exoskeleton requires a high physical effort.The item assessed how much physical activity was required to use the exoskeleton, i.e., whether the task was easy or physically demanding.• I F2 : Using the exoskeleton requires a high mental effort.
The item assessed how much mental and perceptual activity was required to the participant, i.e., whether the task was simple or complex to be performed.

F. Statistical Analysis
To verify the capability of the proposed method to effectively estimate the participants' PP state, a correlation analysis is carried out by comparing the Fuzzy Logic estimations and the self-reported scores.For this correlation, a variation of the PP indicators is calculated as PP = |PP(T f ) − PP(T 0 )|, where T 0 and T f indicate the first and the last time instant of the recording.Considering that participants began Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE III MEAN VALUES AND STANDARD DEVIATIONS OF COLLECTED PHYSIOLOGICAL PARAMETERS PER MFS
Fig. 3.
Physiological responses computed during exoskeletonassisted walking.The shaded areas for each subplot represent the probability density estimation of each physiological parameter j ∈ {HR, RMSSD, LF, RR, SCL, SCR}.Data are colored in green, yellow, and red according to the activation level "LOW", "MID", and "HIGH", respectively.The MFs are superimposed in the same plots to highlight how they can be built up following the approach proposed in Section II-C.

III. RESULTS AND DISCUSSION
Figure 3 shows the distribution of all the physiological responses collected during the experimental sessions.In particular, for each j-th physiological parameter, with j ∈ {HR, RMSSD, LF, RR, SCL, SCR}, the normalized data collected during walking is reported as probability density estimation (in green, yellow, and red for the "LOW", "MID" and "HIGH", respectively).More in detail, Table III resumes the mean values and the standard deviations of the physiological responses obtained per each activation level.
It is worth highlighting that the computed MFs, shown in Figure 3, cover the entire physiological response dataset.The methodology proposed in this paper proved to be an automatic manner to determine the MFs keypoints, thus overcoming the limitation of all the approaches already presented in literature and discussed in Section I. Indeed, both the shape and keypoints of the MFs are typically decided based on researchers' experience, without following a well-defined and standardized procedure.On the contrary, our method allows splitting intrinsically the collected data into subsets representing the activation level of each physiological response.
Particular attention has to be paid to the distribution of the X SCR R data, which shows the highest variability.During the resting phase (RB) the collected SCL exhibits a low intensity activity.Since the X SCR R is computed by normalizing the walking data with respect to the mean value measured during RB, the highly increased spiking activity, as well as the complete absence of spikes, was greatly amplified resulting in a highly ranging physiological response.Indeed, the "MID" MF encodes for all the GSR activity very similar to the one exhibited during the baseline, i.e., the peak of the MF is close to zero (−0.42, as reported in Table III).On the other hand, the "LOW" and "HIGH" MFs represent the lower and higher galvanic activity with respect to the RB.
The time series of the four PP indicators computed based on the data collected for a representative participant (P1) are shown in Figure 4.The PP indicators are averaged on time intervals of 1 min, for a clearer visualization.The temporal evolution of PP indicators over the course of a walking session provides valuable insights into the participants' experience that is a topic particularly relevant in the field of lower limb wearable robotics.Notably, with reference to Figure 4, we found that the evolution of E and S is consistent with expectations, since the progression of the walking session naturally results in an increased physical and mental effort.The validity of this trend is supported by the findings of previous studies reporting similar patterns in related rehabilitation and physical activity contexts [50], [51].Similarly, the increasing pattern of A reflects the improved focus of participants since the effort level intensified.This is in line with findings in [52], [53], which observed an increase in A during strenuous physical activities, thus reinforcing our observations.Furthermore, we noted a gradual decrease in F as the sessions progressed, which can be attributed to a habituation effect.Indeed, participants seemingly reached a more comfortable state during the latter part of the session.This observation is compatible with similar trends found in [54], which also highlighted a decrease in the participants' F acclimated to the exercise conditions.As evident, the Fuzzy Logic method is capable of representing continuous processes like the PP indicators evolution of the participants while walking with the exoskeleton.This capability is quite important for assessing objectively and in real-time how the user is experiencing human-robot interaction since questionnaires are only able to capture the overall experience at the end of the session, possibly with limited objectivity and reliability.
The mean and standard deviation values for the total variation of the PP indicators ( PP) and for the Likert-based scores of the administered items are shown in Figure 5.
In the subjective assessment, the participants state that the robot use is not so physically demanding (I E = 3.4 ± 1.43).The total workload required to carry out the walking session is considered medium (I F = 3.8 ± 1.68).All the participants state that they have to maintain a high level of attention while walking with the robot (I A = 5.6 ± 0.41).Lastly, the stress rating obtained very low scores (I S = 1.7 ± 1.09).
A correlation analysis is carried out to assess the relationship between the estimated objective assessment and the subjective one, as explained in Section II-F. Figure 6 shows the correlation computed between the estimated PP and questionnaire scores.Specifically, each dot represents one participant, the solid line is the regression line, and the shaded area stands for the 95% confidence bounds.It is worth observing that A and F returned ρ = 0.46 (p-value= 0.179) and ρ = 0.60 (p-value= 0.068) and statistical power values of 0.75 and 0.76.Hence, they are considered strongly and moderately correlated with the self-reports, respectively, revealing that the output of the Fuzzy Logic method is consistent with the participants' subjective perception.On the other hand, E and S exhibit a statistically significant very strong inverse correlation with the scores of the questionnaire items.Indeed, the computed correlation coefficients were ρ = −0.62 (p-value= 0.054) and ρ = −0.81(p-value= 0.004), for the E and S, respectively.In this case, the statistical power values are 0.75 and 0.77.
While positive correlations are rather straightforward, negative ones can be explained as reported below.
A high E is obtained for some participants while, on the other hand, the perceived energy expenditure expressed by I E is low.The apparent inconsistency between the questionnaire results and the output of the Fuzzy Logic method could possibly reflect a participant's subjective underestimation of the E. This, in turn, is considered to be ascribed to erroneous participants' initial expectations about their performance.We hypothesize, indeed, that the participants overestimated the needed effort while the final results were actually considered easy to achieve.For this reason, the participants revised their perception of the effort, both physical and mental, and rated it as relatively low in opposition to their initial guess.In summary, the mismatch between the expectation (i.e., a demanding task) and the actual result (i.e., a non-challenging experience) led to a low rate for the I E .
The inverse correlation obtained when comparing S and I S is due two effects.Participants exhibiting a higher modification of the S PP, i.e., whose S level was low at the beginning of the trial (S(T 0 ) < 0.5) and increased during walking according to their physiological modifications, declared minimum perceived frustration.The remaining participants, who had an almost constant S level ≥ 0.5, thus exhibiting small S, stated they felt more uncertain and discouraged during the robot use.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
The proposed Fuzzy Logic method is found to be suitable for computing different nuances of PP indicators based on the variation of physiological parameters.Continuous PP estimation, as opposed to discrete one, can provide a dynamic, real-time perspective of the user's condition.It enables the system to detect PP changes and trigger immediate robot interventions.This adaptability not only improves the quality of rehabilitation but also represents a significant advancement in personalizing and optimizing the rehabilitation experience [55].In particular, the high correlation coefficients |ρ| demonstrate the validity of the method, despite the complexity of the task, and suggests that it can provide sensor-based estimations consistent with the participant's perception of the human-robot interaction experience.The comparison among the estimated PP indicators and self-reported scores, resulting from the administration of the questionnaire, is quite relevant since it is typically not pursued in other literature studies adopting Fuzzy Logic solutions for similar applications [17], [28].Indeed, due to the nature of the estimation under analysis, and the related intrinsic difficulty in defining a reliable objective ground truth, achieving a solid validation of a method for PP assessment is rather challenging.For the same reason, the comparison among different methods employed to analyze PP indicators, and the quality of their results, is nontrivial and cannot be achieved in a strictly rigorous way.Our achievements, supported by a dataset collected on ten healthy participants, can be considered promising.They demonstrate the effectiveness of the proposed method in assessing PP states and the feasibility of applying it to a human-robot interaction scenario.Moreover, they are further supported by the lack of prior validations in similar relevant works.
The investigation pursued in the specific application of exoskeleton-assisted walking is particularly relevant due to a lack of studies in the literature.Different works have been focused on physiological and metabolic responses during walking with exoskeletons, overground [48] or on a treadmill [56], but the analysis of PP state is typically missed and can be considered still an open issue.Our method, preliminary presented in [57], was developed by the Authors, together with a multi-factor questionnaire [13], within the framework of the research project EXPERIENCE (FSTP-1 cascade funding program of the H2020 project EUROBENCH -ICT-2017-1-779963).It is particularly useful because it provides developers of lower limb exoskeletons and clinical users with a tool capable of catching user experience and hence, in line with the vision of the project, benchmarking different exoskeletons, or their control strategies [58].Indeed, the problem of benchmarking robots, particularly those intimately interacting with users, such as wearable ones, has gained a special interest in the robotics community [59].In this context, our method is expected to promote advancements in the knowledge of human-robot interaction features and to support the improvements in the design and usage of exoskeletons.Anyhow, despite the importance of the selected application, it must be highlighted that the proposed approach can be potentially adopted in other operative scenarios to assess the PP state of users during the execution of different physical and/or cognitive tasks.Indeed, since the method is data-driven, it can be easily adapted if dedicated data are acquired under new conditions of interest.

A. Limitations
The experimental results obtained in this study must be interpreted with caution due to the reduced sample size of the enrolled participants.Moreover, our findings are only valid in the case of healthy people and a possible extension to scenarios involving individuals with central nervous system lesions under rehabilitation treatments deserves further investigation.It is also worth highlighting that the analysis carried out in this work compared continuous time-varying PP indicators with questionnaire discrete and single-point scores collected at the end of the walking session.Despite PP indicators provided by our approach assessing what happens in realtime during the activity (even in the presence of additional stimuli and/or environmental conditions), questionnaire scores can only provide a snapshot of the overall experience once the human-robot interaction is finished.Anyhow, the choice of this comparison was forced by the intrinsic difficulty in assessing the user's subjective perspective in real time during the task execution.Lastly, the task under analysis has to be considered not particularly demanding, both physically and cognitively, since the body weight support compensated for the need to manage dynamic balance and also participants were not challenged with specific feedback-based requests.

IV. CONCLUSION
A novel approach for estimating PP indicators from physiological measurements of users executing a motor task has been presented and tested with healthy participants during exoskeleton-assisted treadmill-based walking.It is based on a data-driven Fuzzy Logic method, including custom MFs and rules exploiting literature relationships among PP and physiological variations.
The Fuzzy Logic method has been selected as the most suitable solution to manage highly variable physiological inputs.It has been found to be capable of estimating users' PP state in the selected application.Indeed, experimental results highlighted a high correlation between the estimated PP indicators and self-reported sores resulting from the administration of a dedicated questionnaire.This comparison, typically neglected in similar papers adopting Fuzzy Logic solutions, confirmed the validity of the method and suggests that the PP indicators can be successfully estimated in real-time during the execution of a complex human-robot interaction task.Even if the small number of participants and the partial statistical significance of correlations can represent a limitation for the validity of the study, the presented results are still promising and demonstrate the feasibility of the approach.Nonetheless, the proposed method can be generalized and is suitable to be extended to alternative applications entailing physical tasks and interactions with different robotic agents.Further investigations will be devoted to testing the presented approach on a larger population of healthy participants as well as on people with lesions of the central nervous system.Moreover, future work can be also potentially aimed at applying the same method to evaluate the effect of multiple exoskeleton-assisted walking sessions thus assessing possible adaptation-related longitudinal changes.Finally, another important aspect to be tackled in future works is the analysis of more demanding tasks, including walking with an overground exoskeleton, that intrinsically entails the management of balance and crutches coordination, as well as walking on the treadmill with specific cognitive and physical challenging requests to the users based on visual feedback and verbal instructions.Indeed, the authors are currently working on investigating the effects of different interaction modalities considering specific requests in actively contributing to the motion with positive flexion/extension work on the device.

APPENDIX A RULES FOR THE FUZZY LOGIC METHOD
This appendix includes the fuzzy rules implemented in the method described in Section II-C to transform the physiological measurements X R into the PP indicators of interest.The rules (1-17 for S, 18-34 for A, 35-45 for E, and 46-54 for F) have been defined based on the analysis of the scientific literature analyzing relationships among physiological responses and PP state [11], [37], [46], as follows:

Fig. 1 .
Fig. 1.Functional scheme of the proposed approach to estimate the PP indicators.

Fig. 2 .
Fig. 2. Experimental Setup.A participant equipped with the physiological monitoring system wearing the Lokomat R exoskeleton is shown.
• I A1 : To achieve my results I have to maintain a high level of attention.The item assessed how much hardly the participant had to focus on the use of the exoskeleton and on the task execution.• I A2 : When I walk with the exoskeleton I have to focus a lot on my steps and the movements to be made.The item assessed how much the participant had to focus on physical activity while walking with the exoskeleton.• I S1 : Walking with the exoskeleton is frustrating.The item assessed how much irritated or annoyed the participant felt during the task.• I S2 : When I walk with the exoskeleton I feel uncertain, stressed, and discouraged.The item assessed how much emotionally difficult was to walk with the exoskeleton.For each PP indicator, the mean value of the two items is used in the analysis detailed in Section II-F.
their walking session with a certain PP state, PP expresses the total modification of each indicator after the whole task execution.Moreover, the final time instant is considered relevant because it represents the estimation temporally closest to the moment when the questionnaire is administered to the participant.Since the data are not normally distributed, Spearman's rank correlation coefficient ρ as well as the significance level p-value are computed.The derived correlations are considered very weak if |ρ| ≤ 0.19, weak if 0.20 ≤ |ρ| ≤ 0.39, moderate if 0.40 ≤ |ρ| ≤ 0.59, strong if 0.60 ≤ |ρ| ≤ 0.79 and very strong if 0.80 ≤ |ρ| ≤ 1.0.A correlation is considered statistically significant in the case p-value ≤ 0.05.

Fig. 4 .
Fig. 4. Time series of the four PP indicators computed on the data collected from the participant P1 by using the Fuzzy Logic model.Data are shown at 1 min intervals for the sake of visualization clarity.

Fig. 5 .
Fig. 5.Estimated PP and subjective assessments of the participants.The mean values of PP estimated by the proposed Fuzzy Logic method, computed as the difference between the PP indicator at T f = 15 min and T 0 = 1 min, and the subjective evaluation assessed by the questionnaire are reported.Error bars indicate the standard deviations.

Fig. 6 .
Fig. 6.Correlation computed between each PP and the corresponding questionnaire item I PP .Each point represents one participant.Regression lines are reported together with 95% confidence bounds (shaded areas).Text boxes in the graphs indicate the correlation coefficient ρ and the p-value.

TABLE I DEFINITIONS
OF THE MFS KEYPOINTS FOR THE GENERIC INPUT PHYSIOLOGICAL RESPONSE X j R

TABLE II DEFINITIONS
OF THE MFS KEYPOINTS FOR THE GENERIC OUTPUT PP INDICATOR In (8), a and d locate the feet of the trapezoid, b and c locate the shoulders and x is again the input data.