Exploration of sensations evoked during electrical stimulation of the median nerve at the wrist level

Objective. Nerve rehabilitation following nerve injury or surgery at the wrist level is a lengthy process during which not only peripheral nerves regrow towards receptors and muscles, but also the brain undergoes plastic changes. As a result, at the time when nerves reach their targets, the brain might have already allocated some of the areas within the somatosensory cortex that originally processed hand signals to some other regions of the body. The aim of this study is to show that it is possible to evoke a variety of somatotopic sensations related to the hand while stimulating proximally to the injury, therefore, providing the brain with the relevant inputs from the hand regions affected by the nerve damage. Approach. This study included electrical stimulation of 28 able-bodied participants where an electrode that acted as a cathode was placed above the Median nerve at the wrist level. The parameters of electrical stimulation, amplitude, frequency, and pulse shape, were modulated within predefined ranges to evaluate their influence on the evoked sensations. Main results. Using this methodology, the participants reported a wide variety of somatotopic sensations from the hand regions distal to the stimulation electrode. Significance. Furthermore, to propose an accelerated stimulation tuning procedure that could be implemented in a clinical protocol and/or standalone device for providing meaningful sensations to the somatosensory cortex during nerve regeneration, we trained machine-learning techniques using the gathered data to predict the location/area, naturalness, and sensation type of the evoked sensations following different stimulation patterns.


Introduction
The everyday use of hands is highly dependent on the interaction between receptors and nerves in the peripheral nervous system and the sensory and motor systems in the central nervous system (CNS).The human hand is a highly intricate system using an abundance of information coming from receptors, in the skin, muscles and joints to enable profound interaction with the outside world and to facilitate dexterous control of movement.The complexity and potential of this system are best described by the fact that ∼70 000 sensory axons are dedicated to transferring information gathered by the receptors in each hand to the CNS for further processing and contextualization [1].With each type of sensory receptor reacting to different physical phenomena, such as light touch, temperature, and vibration, and producing action potentials of different temporal patterns [2,3], the CNS is able to synthesize appropriate empirically-based information that we experience as specific sensations [4][5][6].
Following a nerve injury where a nerve is transected and subsequently repaired no afferent nerve signals are sent to the brain leading to physical and psychological impairment [7][8][9].It is estimated that 2%-3% of all the patients admitted to Level I trauma centers have a complete transection of a major nerve [10,11].An injury to a major nerve in the hand most often requires advanced surgical reconstructions using microsurgical techniques.After surgical repair the injured axons have to regenerate from the injury site and reinnervate peripheral receptors, a process usually taking 3-6 months.Unfortunately, some axons die following the injury and some axons reinnervate the wrong receptors.As a result, an adult suffering a major nerve injury in the hand and arm never regains function to the full extent as before the injury.Based on neurobiological prerequisites following a major nerve injury in the arm and hand the rehabilitation process is divided into two phases [9].In phase 1, which starts directly after injury no nerve signals are sent through the injured nerve.This results in cerebral adaptations, i.e. plasticity where the brain area dedicated to processing hand sensations starts processing sensory information from intact nerves in the hand and arm [9,[12][13][14].In phase 1, rehabilitation focuses on stimulating sensory and motor areas in the brain using action observation, sensory and motor imagery, and cross-modal plasticity [15].The main purpose of rehabilitation in phase 1 is to prepare sensory and motor areas in the brain for when the repaired nerve starts sending information again.Phase 2 in the rehabilitation process starts when the patient can perceive some sensibility in the hand or fingers, meaning that the repaired nerve has reinnervated receptors and muscles and started to transmit never impulses.In phase 2, rehabilitation focuses on stimulating the peripheral sensory receptors and muscles.
A common technique used to artificially activate peripheral nerves is electrical stimulation [16].This technique make it possible to directly activate intact nerves as well as nerves that are detached from their target receptors, thereby, providing sensory input to the CNS in a physiological manner.This, in turn, can stimulate the somatosensory cortex in the early phase after a nerve repair [17].Electrical stimulation is based on the use of electronic stimulators producing short voltage or current pulses that are delivered to the body via implantable or transcutaneous electrodes.Even though implantable electrodes provide a possibility to selectively activate groups of nerve fibers within a nerve [18][19][20][21], in the case of rehabilitation where the regaining functionality of nerves is expected through the natural healing process, conducting surgery to deploy extraneural or intraneural electrodes is not justified by the potential long-term gains.Thus, electrical stimulation following nerve repair should be done by transcutaneous electrodes.These electrodes lack the selectivity of stimulation achievable by implanted electrodes but provide an easy-touse interface that can be employed at various stages of the rehabilitation process [22,23].
The main challenge, when using transcutaneous electrodes is to somatotopically match naturallyoccurring sensations.Sensations elicited by electrical stimulation are often perceived as 'unnatural' or of 'low naturalness' [19,24].One explanation for this is that by using (transcutaneous) electrical stimulation it is not possible to match the intricate spatio-temporal patterns of nerve activation seen during hand use in everyday life.Such patterns of nerve activation are resulting from the numerous receptors that register different aspects of handenvironment interaction.In doing so each type of receptor generates a characteristic train of action potentials [25] from which CNS is able to extract qualitative and quantitative properties of hand gestures and interactions.Even with sophisticated methods of recording and analyzing neural signals, we are not able to comprehend the neural coding to mimic it via a neural interface to produce rich sensations artificially.
Numerous attempts at decoding naturallyoccurring sensations and simulating natural-like sensations through the use of electrical stimulation have been reported in recent years.By using invasive techniques it is possible to target a subset of nerve fibers within a nerve, leading to moderate success in producing several different somatotopic sensations on the skin [18,19,21,26].The type of sensation and the active area are usually limited by the positions of the electrodes which, in the case of implanted electrodes, can not be dynamically changed to extend the somatotopic activations.Moreover, even slight changes in electrode position, tissue conductive properties, and generated stimulus properties significantly affect sensation type, active area, and naturalness perceived.In the case of transcutaneous stimulation, it is even harder to achieve reliable and robust somatotopic sensations [27].
The aim of this study was to evaluate sensations in the hand resulting from transcutaneous electrical stimulation of the Median nerve at the wrist.Using the data gathered during this experiment, a machinelearning algorithm was employed to generalize the obtained results and provide associations between stimulation parameters (frequency, amplitude, and shape) and the type, area, and naturalness of the perceived sensations were assessed.This approach of using a machine-learning model to predict the elicited sensation also presents the main novelty of the paper.

Subjects
28 able-bodied participants (23.4 ± 6.7 years old), 16 women and 12 men, participated in the study.Inclusion criteria were: age between 18 and 50 years.Exclusion criteria were: prior surgery to the hand, neurological or psychiatric disorder, prior frequent exposure to hand-held vibrating tools, diabetes, pregnancy, an implanted pacemaker, or heart monitoring device.None of the participants had any previous experience with electrical stimulation.All participants provided informed consent, and the study was approved by the Swedish Ethical Review Authority (DNR 2020-03937).

Equipment
A custom-made electronic stimulator capable of producing biphasic charge-balanced triphasic (prepulse, cathodic pulse, and charge-balancing pulse) currentcontrolled pulses of amplitudes in the range from 0.1 mA to 10 mA, with a gradual increase in steps of 0.1 mA, and frequencies of 1-100 Hz was used.The stimulation was routed to two self-adhesive Pals transcutaneous electrodes (Axelgaard Manufacturing Co., Lystrup, Denmark).The cathode (round 2.5 cm in diameter) was placed over the Median nerve of the left hand of the participants, in between the tendons from the flexor carpi radialis and palmaris longus muscles, at the wrist level, while the anode (oval 4 × 6 cm) was placed on the volar side of the forearm just distal to the antecubital fossa.Using a significantly larger electrode surface of the anode electrode (24 cm 2 ) compared with the cathode (4.9 cm 2 ) results in a smaller current density under the cathode, consequently reducing the chance of activating nerves underneath it during the pre-pulse or compensation pulse in the stimulus complex.
The stimulation parameters were set and triggered by a custom-made LabView (National Instruments Corp, Austin, Texas, US) program that was made in the form of a graphical user interface (GUI), similar to [28].The program automatically generated a random sequence of stimulation trains with all permutations of the selected stimulation parameters.Following a stimulation train lasting 2 s, the participants were able to fill in information related to the experienced sensation in GUI using their right hand.The questionnaire implemented in the GUI was derived from the paper published by Kim et al [29].The information required by the questionnaire was: 1. 'Location'-markings on the specific hand areas where stimulation was felt.2. 'Naturalness'-integer scale 1-5 where 1 depicts unnatural and 5 completely natural sensations 3. 'Type of sensation'-by selecting one of the 8 options: no sensation, tingle, buzz, pulse, tap, twitch, prick, and 'other' Upon selecting 'Type of sensation' which was located rightmost in the GUI, the program automatically generates the next stimulation pattern.The time dedicated to the interaction with GUI was not limited.Furthermore, before deciding on the 'Type of sensation' , the participant was able, if needed, to repeat the last stimulation pattern an unlimited number of times.Consequently, the period between two different stimulation patterns was governed by the participant, and on average was between 30 s and 60 s.

Protocol
At the beginning of each session, the experimenter explained the goals of the study, the recording protocol, possible risks, and the equipment to the participant both orally and in writing, after which he/she confirmed willingness to participate in the study and signed the Informed consent.The main focus was placed on describing different sensations and naturalness using the questionnaire derived from Kim et al [29] so that the results could be consistent within the subject group, and also be compared with similar studies using the same questionnaire.
The participant was seated comfortably in a chair.The non-dominant forearm and hand were placed in a supinated fashion on the table in front of them where it was supported by a cushion.Electrodes were placed on the forearm as shown in figure 1(A).The dominant hand was used to interact with the GUI.
To estimate the lower and upper limits of electrical stimulation, an initial stimulation procedure was performed for each participant.During this procedure, the experimenter set the stimulus width to 250 µs, stimulation frequency to 50 Hz, stimulation duration to 2 s (100 pulses), and stimulus amplitude to 1 mA.From this set point, the stimulation amplitude was increased by 0.1 mA, while keeping other parameters constant, until the participant was able to detect a 2 s pulse train.This stimulation amplitude was then used as the sensation threshold for that particular participant.In accordance with the expanded and revised version of the Short-form McGill Pain Questionnaire [30], this threshold was used as level 1 on the sensation scale.The upper limit of the stimulation amplitude was defined as level 8 on the sensation scale, which was described to the participants as pain that could not be tolerated for more than 20 s.To find the pain threshold, the stimulation amplitude was increased in 1 mA increments to reduce the setup time and inconvenience for the participant.
Based on the established upper and lower sensation thresholds, the ranges for stimulation parameters were set for each participant individually as follows: −Current (level 7) ], by assuming a linear relationship between the stimulation amplitude and pain level.Although the cathodic pulse is always negative, the amplitude is commonly expressed as the positive value, as in [31].This convention was used within this paper too.• Cathodic pulse width: 250 µs • Pre-pulse width: 500 µs • Pre-pulse amplitude (PPA): [−Current (level 1 )/2 + 0.1 mA−Current (level 1 )/2-0.1 mA].After the division, the result was rounded down to the nearest 0.1 mA value.Similar to the value convention used for the cathodic pulse, positive PPA To assess the influence of pulse shape on the elicited sensations a subthreshold pre-pulse was adjoined to the biphasic stimulation pulse, constructing a triphasic pulse.The application of pre-pulses to alter the cathodic pulse follows simulation studies reporting the consequence of using hyperpolarizing and depolarizing pre-pulses on the generation of action potential in nerves [31,32].This effect is based on the non-linear properties of the neuron membrane, where the subthreshold pre-pulses modify the resting membrane potential, therefore changing both the amplitude of the cathodic pulse necessary to produce an action potential, but also the volume within stimulated tissue in which the nerve fibers will be activated.
The rationale for using pre-pulses within the specified range was to modulate the stimulus shape by adding a subthreshold pulse, which, on its own, was not able to result in nerve activation.In addition, the upper limit of CPA was reduced to the amplitude associated with level 7 sensation as the positive pre-pulse was increasing the total cathodic current.However, due to the short duration of stimulation trains and the habituation of the participants to the electrical stimulation, the sensations during the measurement were very rarely perceived as painful.
To limit the time of the measurement to less than 1 h, the ranges for CPA, PPA, and stimulation frequency were divided into six, three, and five levels, respectively, resulting in 90 permutations.In the case of PPA which had three levels, the middle level corresponded to 0 mA (only the main cathodic pulse was produced) while the lowest and highest levels were symmetrical to 0 mA.The permutation of the stimulation parameter was done automatically for each participant.

Data evaluation
The data were stored in text files containing a table of stimulation parameters and responses entered by the participant using GUI.The basic data analysis focused on: (1)extracting locations of elicited sensations, (2) naturalnesses, and (3) occurrences of different sensations from all subjects.The more in-depth analysis of the data was focused on: 1. Assessing how stimulation parameters correspond to the type of stimulation, the naturalness of stimulation, and the area of sensation 2. Devising and evaluating a method for predicting the type of stimulation, the naturalness of stimulation, and the area of sensation upon delivering a stimulation train.
In addition to varying the stimulation parameters (CPA, PPA, and stimulation frequency), total pulse amplitude (TPA) was defined as CPA + PPA to evaluate the correlation between electric charge (cathodic pulse) and the elicited sensations.To find the association between stimulation patterns and the area of sensation (sum of locations on the hand where sensation was felt) Pearson cross-correlation test was used.The same test (Pearson cross-correlation) was also used to find the association between stimulation patterns and naturalness as the naturalness was considered a numeric and continuous variable.In the case of sensation types which are categoric variables, 'ANOVA feature importance scores' were used to associate sensation types and stimulation parameters.
By using stimulation parameters as input variables three decision tree models with 100 splits were constructed to predict/classify: (1) the area of sensations, (2) the naturalness of stimulation, and (3) the type of sensation.The recorded data were randomly divided between training and testing sets in an 80%-20% ratio.The process of randomly dividing the data and training/testing decision tree models was repeated 1000 times to obtain average ± std accuracies of the models.
As the 'area of sensation' and 'naturalness' are discrete variables, where e.g.naturalness of 3 is better than naturalness of 2 and worse than naturalness of 4, some of the prediction/classification errors are less severe than the others.Namely, predicting the high naturalness of the stimulation pattern (e.g.5), while actually producing a completely unnatural sensation (e.g. 1 on the naturalness scale), is an error of larger consequence than mispredicting neighboring naturalnesses.Therefore, in addition to assessing the accuracy of the classification, we evaluated the performance of the decision tree model using Cohen´s kappa.The main characteristic of the Cohen's kappa method is that the entries in the confusion matrix after the classification are weighted so that the values close to the main diagonal are contributing less to the error of the classifier, while values away from the main diagonal are penalized more.From the Cohen's kappa method, two variables were taken as indicators of the capabilities of the decision tree classifier: Observed agreement and Cohen's kappa coefficient.The Observed agreement value was calculated over the confusion matrix using linear weighted coefficients, resulting in a percentage between 0 and 100%.The calculated Cohen's kappa coefficient was interpreted in ranges between 0 and 1 that correspond to different agreement levels between ground truth (reported value) and value estimated by the decision tree models.
As data did not pass the normality test (Lilliefors test) in all of the analyses, the statistical significance (p < 0.05) between different groups of parameters was assessed using the Kruska-Willis test with Bonferroni post hoc correction for multiple comparisons.All the data post-processing was done in Matlab2022a (MathWorks, Natick, MA, USA).

Results
Electrical stimulation of the Median nerve at the wrist elicited sensations in the hand, as shown in figure 2. The majority of the sensation was felt in the middle finger and palm, areas associated with the Median nerve.However, there were some sensations felt in regions associated with the ulnar nerve, e.g. the subregion marked by i.No participant reported sensations on the dorsal side of the hand, which is understandable as it is innervated mostly by the Radial nerve branches.

Sensation area
The sensation area was defined for each stimulation pattern as the number of hand sub-regions selected by the participants.In an example, selecting sub-regions associated with the index and middle finger (d, e, f, and g) would result in an area of size 4. Figure 3 shows the accumulative results of the sensation area for all participants and all stimulation patterns, excluding the cases when 'No sensation' was selected (sensation area = 0).
Pearson cross-correlation test was used to assess the contribution of each stimulation parameter to the sensation area.As can be seen in table 1, all parameters were correlated with the sensation area (p < 0.05), where CPA and TPA had the largest correlation coefficients.
The results of applying different TPA on the sensation are shown in figure 4. As expected, increasing stimulation amplitude (in this case TPA) leads to an increase in the hand area where sensations were felt.The statistical analysis between sensation areas resulting from different TPA-s showed that there are four groups of TPA-s that elicit significantly different sensation areas.TPA-s between levels 1 and 4 (level 2 is the level just above the sensory threshold found before initiating the protocol with randomized stimulation patterns) most frequently activated 2-3 subregions of the hand (two to three sub-regions in the GUI were selected), TPA-s between 5 and 6 activated 4 sub-regions, TPA of 7 (just below pain threshold found before initiating the protocol with randomized stimulation patterns) activated 5 sub-regions, while TPA of 8 most frequently activated 7 sub-regions of the hand.
In an attempt to predict the area of sensation for a stimulation pattern, a decision tree model was trained and tested on the dataset.Following 1000 simulation runs with different splits between training and testing data, the accuracy of predicting specific areas by the decision tree was 0.15 ± 0.02 (mean ± std), see figure 5(A), which is double the chance level (1/13).
By analyzing errors of the decision tree classifier (see figure 5(B)), it should be noted that the majority of the misclassifications were close to the main diagonal.Therefore, when calculating Cohen's   kappa, the Observed agreement between the subjectselected area and the area inferred by the decision tree classifier was 0.81 ± 0.01 (mean ± std).Cohen's kappa coefficient was 0.15 ± 0.03 which is interpreted as 'slight agreement' between ground truth and the inferred value.

Naturalness
The cumulative naturalness for each participant and each stimulation pattern is shown in figure 6.
Similarly to the analysis of areas where sensations were felt, the Pearson cross-correlation coefficients were calculated between stimulation parameters and experience of naturalness, see table 2. This analysis revealed a high (negative) correlation between pulse amplitude, primarily TPA.On the other hand, stimulation frequency was not significantly correlated to the naturalness (p > 0.05).
The influence of different levels of CPA on the naturalness is shown in figure 7(A).As the crosscorrelation coefficient indicates, increasing the CPA gradually decreased how natural the sensations were experienced.The statistical analysis of the elicited naturalness resulting from different CPA-s showed no statistically significant difference between adjacent levels of CPA, e.g. level 4 CPA is not significantly different from levels 3 and 5, but produced a different feeling of naturalness compared to other levels of CPA. Figure 7(B) shows the results of different prepulse shapes.There is no statistically significant difference between negative pre-pulse and no pre-pulse, but there is a statistically significant drop in the naturaleness when using just below-threshold cathodic pre-pulses.
Following 1000 simulation runs with different splits between training and testing data, the accuracy of predicting naturalness by the decision tree was 0.33 ± 0.02 (mean ± std), where the chance level is 0.2 (1/5).Similarly to the Sensation area, the confusion matrix for the Naturalness showed that the majority of misclassifications happen close to the main diagonal, see figure 8(A), resulting in a relatively high Observed agreement metric of 0.75 ± 0.01.Cohen's kappa coefficient was 0.21 ± 0.03 (see figure 8(B)) which is interpreted as 'fair agreement' between ground truth and the inferred value.

Sensations
Analysis of sensations was done using a different methodology than when analyzing sensation area and naturalness as the sensations are categorical variables.The variety of sensations elicited while stimulating    the Median nerve is shown in figure 9. Except for 'No sensation' which was selected for most of the low TPA-s, the most common sensations were 'Buzz' which was experienced following 26% of stimulation patterns, and 'Tingle' which was experienced following 18% of stimulation patterns.Nevertheless, from figure 9 it can be noted that all sensations were represented with a fair rate of occurrence.In 0.2% of cases, the participants could not associate the elicited sensation with any of the options in the GUI, thus, they selected the 'Other' choice.The importance of stimulation parameters for eliciting specific sensations was calculated using a feature importance score based on ANOVA.The negative logarithm of p-values was used to determine the underlying connection between parameters and sensations.As presented in figure 10, the most dominant factor for eliciting sensations was the  frequency of electrical stimulation delivered through the electrodes (importance score of 295), followed by TTP (importance score of 177).
As the stimulation frequency is the most important parameter for eliciting specific sensations, figure 11(A) shows the most common connections between stimulation frequency and sensation.As it can be noted from the figure, tingle, buzz, and prick sensations usually result from the stimulations of high frequencies, while pulse, tap, and twitch can be evoked by electrical stimulation using low-frequency pulse repetition.Although there is an overlap between the sensations for the same range of stimulation frequencies, other parameters also play significant roles in forming specific sensations (see figure 11(B) on how pre-pulses influence elicited sensations), which enables a machine-learning algorithm to find most probable combinations of stimulation patterns that lead to eliciting specific sensations.
In an attempt to predict the exact sensation type based on the stimulation pattern, 1000 simulation runs with different splits between training and testing data using the decision tree method were executed.The accuracy of classification was 0.5 ± 0.02 (mean ± std), see figure 12(A), which is well above the chance level of 0.17 (1/6).As the set of elicited sensations was unbalanced, Tingle and Buzz sensations were overrepresented compared to other sensations.Therefore, balanced accuracy was also calculated.It is defined as the mean of sensitivity and specificity scores for each class and had a value of 0.32 ± 0.01.
The normalized confusion matrix is shown in figure 12(B).As evident in figure 11(A), there is an overlap between the stimulation frequencies resulting in tingle and buzz sensations, which is reflected in the errors in separating these two sensations by the decision tree classifier.Due to the unbalanced set of data, the decision tree seemed to favor buzz sensation.This is especially evident in the case of twitch and prick sensations that in 80%-20% division between training and testing set were sometimes significantly underrepresented in the training data, resulting in classification/prediction errors where most of the predicted sensations were assigned to either tingle or buzz.

Discussion
This study shows that transcutaneous electrical stimulation on the Median nerve at the wrist can evoke sensations from the Median nerve innervated areas in the hand.Furthermore, by modulating the stimulation parameters such as frequency and amplitude of the pre-pulse and the cathodic (main) pulse, effectively altering the pulse shape it was possible to elicit different sensations, how natural these sensations were experienced, and where in the hand they were experienced.Collecting the data on 28 participants enabled the use of machine-learning to generalize causality between different stimulus shapes and elicited sensations.This presents the main contribution of this paper as it shows that the stimulation parameters can be pre-tuned for specific sensations, therefore reducing the setup procedure in clinical applications that require eliciting certain somatotopic sensations using electrical stimulation.
Artificially eliciting somatotopically matched sensations is a desirable feature in many applications within the field of neuroengineering.Providing a natural-like sensation is beneficial for amputees using prosthetic devices [33,34], persons undergoing rehabilitation following peripheral nerve injury [17], stroke [35,36], or being affected by neurodegenerative disorders [37].
The first part of the analysis focused on the evaluation of the sensation area following a pattern of electrical stimulation.As shown in figure 2, the sensations were elicited in all the hand regions innervated by the median nerve, which was the primary target of this study.The sensations were felt in all parts of the hand innervated by the median nerve although most frequently felt in the middle finger  and thenar area.The incidence of sensations felt in different regions (figure 2) is in agreement with the topology of the median nerve fascicles in the forearm [38,39].Planitzer et al [38] provided evidence that the parts of the median nerve closest to the skin (the palmar side of the median nerve) most commonly consist of fascicles going from the palmar cutaneous branch (89% of fascicles belonging to this branch are in the nerve volume closest to the skin), the middle finger brunch (>60% fascicles) and branch belonging to the radial side of the ring finger (>70% fascicles), which are the parts of the hand were also most frequently marked during our study.Furthermore, a similar observation regarding the internal structure of the median nerve is evident in patients with compressions of the median nerve at the wrist, commonly known as carpal tunnel syndrome.Clinically patients with carpal tunnel syndrome most often experience numbness in the thenar and the middle finger, indicating that these nerve fascicles are located closest to the carpal ligament [40].Interestingly, a small number of stimulations resulted in sensations felt in the ulnar nerve innervated skin areas.There are several possible explanations for this result.Due to the proximity between the Median and ulnar nerve at the wrist in the distal forearm, especially in persons with a narrow/smaller arm, when a higher pulse amplitude is delivered transcutaneously to the area of the median nerve at the wrist, it is likely to activate also the ulnar nerve.Similarly, a slightly misplaced electrode, even for a few millimeters, can result in co-stimulation of the ulnar nerve.Another explanation can be based on the anatomical variations between participants.The ring finger is innervated by the median nerve (the radial part of the finger) and the ulnar nerve (the ulnar part of the finger).However, it is known that the innervation of the ring finger is subject to very large anatomical variation.For example, it is not uncommon for patients to feel a light touch on the ulnar side of the ring finger even if they have a complete transection of the ulnar nerve.Thus, stimulation of the median nerve in the distal part of the forearm can very well result in sensory experiences from the entire ring finger.
The study shows that by changing the stimulation parameters it is possible to affect/control the size of the area in the hand the subject experiences sensations.As shown in figure 3, most frequently, only 2-3 hand sub-regions were indicated as the sensation area, but there were cases when a stimulation pattern activated the entire volar side of the hand.There was a statistically significant cross-correlation between all stimulation parameters and the sensation area, where TPA had the highest correlation coefficient of 0.36.Further analysis of the interaction between TPA and sensation area showed that it is possible to divide TPA into four groups, depending on how many hand sub-regions are activated.For low levels of TPA (1-4) we can expect that the most frequently 2-3 hand sub-regions are activated, followed by the TPA levels 5-6 most frequently resulted in activation of 4 hand sub-regions, the TPA level 7 activated 5 hand sub-regions and the TPA level 8 activated 7 hand sub-regions.This aligns with the theory where progressive activation of larger hand areas is expected following the increase of the stimulation current amplitude [41].
The Naturalness of the elicited sensations was also reported in the full range of proposed values, with levels 2-4 being more frequently used to describe the felt sensation, see figure 6.The statistical analysis showed that there are significant negative correlations between pulse shape (pre-pulse amplitude, cathodic pulse amplitude, and TPA) and the naturalness of the sensation, but not between frequency and the naturalness, where the highest correlation coefficient was for TPA (−0.37).This finding provides evidence that the stimulation with amplitude just above the sensation threshold yields sensations of the highest naturalness.Achieving such optimal stimulation amplitude carries a risk of a gradual shift of the sensation threshold due to habitation, which can result in the stimulation being not perceived.Furthermore, lower (total) stimulation amplitude is associated with the sensation localized to a smaller hand area, therefore there is an implicit trade-off between the naturalness and the area of the felt sensations, which is in line with the literature showing similar findings when stimulated using intraneural electrodes [19].The evaluation of the effect of using pre-pulses revealed that it is not beneficial to include cathodic pre-pulses as they yield a statistically significant reduction of the naturalness of the elicited sensations.Anodic pre-pulses did not show negative effects on the naturalness of the sensation, and they might be effective for tuning the type of elicited stimulation as there is a statistically significant impact on that aspect of the somatotopic sensation.
The interaction between one of the significantly correlated pulse shape variables, CPA, is shown in figure 7. The data presented in this figure imply that there are no significant differences among the neighboring CPA levels, therefore indicating that to increase the naturalness of felt sensation, the CPA should be reduced by at least 2 levels/steps.
The sensations that were elicited were constrained in GUI into eight options including 'No sensation' and 'Other' .As figure 9 shows, it was possible to achieve each of the selected sensations using different stimulation patterns.In the case of targeting different sensations, the main contribution comes from the stimulation frequency, followed by the TPA.This shows the complementary nature of different stimulation patterns as for the sensation area and naturalness, stimulation frequency was not highly (or not at all in the case of naturalness) correlated with these descriptors of the elicited sensations.The mechanism of translating transcutaneous electrical stimulation into somatotopic sensations is a complex scientific question that includes several steps, (1) how a sensation is coded by the receptors, (which afferent neurons are more likely to be activated following external electrical field, and (3) how CNS filters and interprets synchronous action potentials coming from the hand area.While there are (forward) models describing the behavior of hand receptors [42] and insights on how to selectively activate nerve fascicles using implanted electrodes [43], it is not clear how synchronous activation of relatively large portions of a nerve with fascicles belonging to different types of receptors, each naturally firing at different frequencies, is processed by the CNS.On the other hand, here we present a very practical approach that uses a data-driven model to predict sensations generated by electrical stimulation to accelerate the setup process.
The current study also explored the possibility to predict elicited sensation descriptors based on the stimulation pattern using a machine-learning technique.To the best of our knowledge, such an approach has not been explored before.In most prior studies, the methods for tuning the stimulation parameters to achieve the desired sensation rely on varying stimulation parameters in a broad range of values while the subject reports the elicited sensation.Although there are some statistical regularities of how the median nerve is spatially divided into fascicles of branches going from different hand areas, the common approach is based on the notion that the elicited sensations following electrical stimulation are highly subjective as the excitability, position, and internal structure of the nerves are unique for every person, therefore requiring calibration for each person and each stimulation session.There are studies that investigated methods for accelerated tuning the electrical stimulation which are based on different convergency techniques [44], but the whole process still typically depends on exploring stimulation parameters within a predefined, usually wide, range.In the current study, data was gathered from 28 participants, where each characterized 90 sensations resulting from different stimulation patterns.This amount of data made it possible to use machine-learning techniques to generalize the association between stimulation parameters and elicited sensations.For this purpose, a decision tree classifier was used.The basic results of the classification showed that it is possible to predict, even with this relatively simple machinelearning model, the outcomes of stimulation patterns with accuracy well above the chance level for all three descriptors of elicited sensations: (1) sensation area: 0.15 ± 0.02 (chance level 0.077), (2) naturalness: 0.33 ± 0.02 (chance level 0.2), and (3) Sensation type: 0.5 ± 0.02 (chance level 0.17).As the sensation area and naturalness are discrete numerical variables subjectively chosen by the participants, the classification accuracy provides only partial information regarding the classifier performance in a real-world application.Therefore we computed additional metrics based on the Cohen's kappa method to evaluate the impact of classification errors on the predicted sensation.Due to the fact that most of the misclassifications occur at the neighboring values (close to the main diagonal of the confusion matrix), the observed agreement metric is relatively high: (1) sensation area: 0.81 ± 0.01, and (2) naturalness: 0.75 ± 0.01, indicating that in most of the cases, the predicted value would fall relatively close to the 'true' descriptor of the elicited sensation.Therefore, based on these findings, we suggest that the use of a machine-learning model can effectively adjust the stimulation parameters for each individual and threby provide specific sensations during nerve rehabilitation.This can provide the somatosensory cortex with meaningful sensory inputs in phase 1 following nerve repair and thus better prepare the brain for phase 2 in the rehabilitation process.
When predicting sensation type using machinelearning, the classes were highly imbalanced which was also reflected in the classifier outputs that favored overrepresented sensations.Therefore, to also provide an understanding of the real-world scenario where the underrepresented sensations would be equally 'targeted' as the overrepresented ones, we calculated the balanced accuracy which, using our dataset, was 0.32 ± 0.01.Although the balanced accuracy of predicting the sensation type is relatively low, it has to be noted that although relatively large, our dataset was still limited in size in terms of machine-learning standards.
The use of machine-learning to predict descriptors of the sensations generated by transcutaneous electrical stimulation could not be compared with any previous studies.In the context of some other fields of neural engineering, such as prosthetic control, the classification accuracies presented in this paper fall far below the state-of-the-art myocontrolbased classification systems [45].On the other hand, compared with the current methodology of tuning electrical stimulation to provide somatotopic feedback which is commonly based on 'grid search' approaches with a wide range of values [21,46], the possibility to limit the test range using suggestions from a machine-learning algorithm, would greatly reduce setup and tuning time of a system intended for stimulating major nerves.
During the measurement, the placement of the electrodes was guided by the physiological landmarks of the forearm.Nevertheless, we expect that there was a certain level of variability of electrode placement between the participants, leading to the activation of different nerve portions during the stimulation, i.e. more radial or ulnar part of the median nerve.In addition, as shown in Planitzer et al [38] there is natural interperson variability in the topology of the median nerve fascicles in the forearm.To minimize the impact of extreme differences in electrode placement and the median nerve topology, a relatively large sample was used (28 subjects) while relying on machine-learning to generalize the results and provide the most common connections between the stimulation parameters and the evoked sensations.
The limitation of the present study is mainly related to the highly subjective criteria for the Sensation type and the Naturalness that were reported by the participants.Although the study protocol included detailed instructions based on the descriptions provided in [29], it still posed a problem for the participants to select the most appropriate sensation and naturalness, that matched the elicited sensation, from the list.This might have led to inaccuracies in our findings and classifications due to the different interpretations and self-conceived criteria of how to describe a felt sensation.In addition, naturalness was defined as a scale of five discrete values ranging from 'totally natural' to 'totally unnatural' , where the participants were asked to estimate the likelihood of the elicited sensation happening in everyday life.As we were not able to provide examples of sensations generated by electrical stimulation for different naturalness levels, it was up to the participants to establish the subjective scale.In addition, it is possible that during the measurement, the self-imposed criterion of rating the Naturalness of the sensations was changing as the participants experienced new sensations.This might influence the consistency of the gathered data on both, intersubject and intrasubject levels.

Conclusion
This study assesses how transcutaneous electrical stimulation on the Median nerve at the wrist is experienced.The stimulation parameters that were modulated were the stimulation frequency and amplitudes of the pre-pulse and the cathodic (main) pulse, effectively altering the pulse shape.Results from 28 ablebodied participants showed that it was possible to elicit the full range of the sensations included in the subset commonly used in studies exploring electrotactile feedback.The naturalness and the sensation area of the felt sensations were also changing in the wide range while modulating stimulation parameters.Furthermore, using the obtained database we assessed some of the underlying connections between the stimulation parameters and how the resulting sensations were described in terms of sensation type, naturalness, and sensation area.The application of the machine-learning (decision tree classifier) method to predict the resulting sensations showed promising results that could be used in future studies to reduce the time needed to tune the stimulation parameters for a specific somatotopic sensation.

Figure 1 .
Figure 1.(A) Electrode setup.The larger electrode acting as an anode was placed on the volar side of the forearm close to the antecubital fossa, while the cathode was positioned above the median nerve at the wrist.(B) GUI that was used to obtain information regarding the elicited sensations.(C) Depiction of pulse trains, each lasting for 2 s, separated by the time needed for the participant to describe the elicited sensation through the interaction with the GUI.

Figure 2 .
Figure 2. Occurance of elicited sensations in different sub-regions of the hand.

Figure 3 .
Figure 3.The area (cumulative number of selected sub-regions of the hand) of sensations elicited during measurements.

Figure 4 .
Figure 4. Boxplot of sensation area as the result of using different TPA.The numbers with asterisk indicates groups of TPA-s that are significantly different from each other (p < 0.05), but not within the group.The comparison was done using the Kruska-Willis test with Bonferroni post hoc correction.

Figure 6 .
Figure 6. of different naturalnesses during the study.

Table 2 .Figure 7 .
Figure 7. Boxplots of naturalness of elicited sensations as the result of using different (A) CPA and (B) PPA levels.The asterisk indicates statistically significant differences between levels of CPA (p < 0.05).The comparison was done using the Kruska-Willis test with Bonferroni post hoc correction.

Figure 8 .
Figure 8. (A) Normalized confusion matrix depicting the ability of the decision tree classifier to predict the naturalness of the elicited sensation.(B) Cohen's kappa coefficient over 1000 simulation runs.

Figure 9 .
Figure 9.The sensations elicited during the study.

Figure 10 .
Figure 10.The feature importance score based on ANOVA (filter method for feature selection).

Figure 11 .
Figure 11.(A) Boxplot of stimulation frequencies eliciting specific sensations.(B) Boxplot of PPA eliciting specific sensations.The asterisk indicates statistically significant differences between stimulation frequencies (p < 0.05).The comparison was done using the Kruska-Willis test with Bonferroni post hoc correction.

Figure 12 .
Figure 12. (A) Classification accuracy over 1000 simulation runs.(B) Normalized confusion matrix depicting the ability of the decision tree classifier to predict the type of elicited sensations.

Table 1 .
Pearson cross-correlation coefficients between sensation area and stimulation parameters.All cross-correlations are statistically significant (p < 0.05).