This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/

Novel communication techniques have always been fascinating for humankind. This pilot study presents an approach to human interaction by combining direct brain-to-brain interface (BBI) and muscle-to-muscle interface (MMI) in a closed-loop pattern. In this system, artificial paths (data flows) functionally connect natural paths (nerves). The intention from one subject (sender) is recognized using electroencephalography (EEG) based brain-computer interface (BCI), which is sent out to trigger transcranial magnetic stimulation (TMS) on the other subject (receiver) and induce hand motion; meanwhile TMS results in a significant change on the motor evoked potentials (MEP) recorded by electromyography (EMG) of the receiver’s arm, which triggers functional electrical stimulation (FES) applied to the sender’s arm and generates hand motion. Human-controlled loop and automatic control loop experiments were performed with 6 pairs of healthy subjects to evaluate the performance of the introduced mechanism. The results indicated that response accuracy during human-controlled experiments was 85% which demonstrates the feasibility of the proposed method. During the automatic control test, two subjects could accomplish repetitive and reciprocal hand motion control up to 85 times consecutively.

In addition, an engrossing artificial communication technique is used for building an information path between two muscles, we call it muscle-to-muscle interface (MMI) herein. In the rehabilitation field, MMI is generally introduced by using electromyography (EMG) for functional electrical stimulation (FES) control 7 . In previous studies, EMG-controlled FES was mostly applied to muscles of one and the same subject for rehabilitation 8 . Recently, EMG-controlled FES was used in a master-slave paradigm between two persons, indicating as well that MMI can be a feasible information transfer approach for human-to-human control 2 . Despite the fact that BBI and MMI technologies have separately substantiated as novel ways to establish artificial communication between two functioning organisms, the result when using them in a unified mechanism remains obscure. This work aims to develop closed-loop control between two persons based on BBI and MMI as shown in Fig. . We expected to build two artificial pathways that functionally connect two natural neural pathways as a closed information loop. We adopted EEG-based BCI and TMS to construct one artificial pathway (i.e. BBI), and EMG-triggered FES to form the other one (i.e. MMI). FES can evoke EEG through afferent nerves in one person, and TMS can induce EMG through efferent nerves in the other person. The realization, as well as the performance of the current system, is presented in this paper.

The conventional interactions between two humans or animals basically depend on vision, audition, voice, olfaction or touch. However, new technologies, such as brain-to-brain interface (BBI) and muscle-to-muscle interface (MMI), have been proposed based on unconventional approaches to explore the novel concept of interactive communication 1 , 2 . BBI, which emerged as an extension of brain-computer interface (BCI), aims to transfer information between two individuals merely using their brains without any intentional physical motion. This technique was first tested on communication among two brains of a pair of functioning rats to jointly learn and move in synchrony. In this study M1 neural ensemble was used as a motor information elicitation source in the encoder rat and invasive intercortical microstimulation (ICMS) as corresponding command inducer in the decoder rat’s brain 3 . BBI was also used to successfully establish artificial information transfer pathway from a human to a rat using electroencephalography (EEG) and transcranial focused ultrasound (FUS) to control a simple motion of the rat’s tail using BBI 4 . In a later study, the first direct brain-to-brain interface between two humans was established, which investigated the feasibility of decoding a command from a sender’s brain by EEG and forcing a receiver to follow the command using transcranial magnetic stimulation (TMS) 1 . Analogously, internet-based human brain-to-brain communication also succeeded in implementing conscious word transmission, over a long distance, exploiting EEG and TMS by means of Bacon’s cipher 5 . BBI has inspired other interesting application, by adopting steady-state visual evoked potential (SSVEP)-based BCI for the human and applying invasive neural stimulation to a cyborg cockroach, the human could control the cockroach to walk along a trajectory 6 . Although some achievements have been accomplished through establishing functional BBIs, all the previous work just focused on single-way communication and realized open-loop control between two subjects, i.e. only the sender can control the receiver.

Mutual information 11 , as an efficiency indicator of information transferred between two subjects, is carried out to evaluate the performance of the current system. The amount of mutual information transferred is calculated by multiplying response vector I(A, B) with the number of related trials. Here, the whole system is disassembled into two parts. In the first part, we obtained the mutual information transferred from the subject’s brain (side A) to the other subject’s muscle (side B), as illustrated in Table . The results demonstrate that the average information transferred via brain to brain to muscle varies between 18.82 ± 7.99 bits for pair 3 and 26.62 ± 4.432 bits for pair 5. Moreover, for the second part, mutual information transferred from the subject’s muscle (side B) back to the subject’s brain (side A) are determined and shown in Table . According to this table, information transferred in the remaining chain via brain to brain to muscle vary from 9.22 ± 1.27 bits for pair 1 to 10.82 ± 2.73 bits for pair 5. Even though there is common inter-subject variability for BCI dependent step, remaining steps are perceived to work reliably. In previous results, it is indicated that MMI shows a better performance than BBI, however, less mutual information is transferred through MMI herein. This occurrence is basically due to sequential nature of the mechanism which allows only those trials with a prosperous BBI to reach MMI section, which means half of the trials are related to the second part. Thus, with respect to the linear correlation between the transferable amount of mutual information and number of trials, less information can be transferred in MMI to brain section. Accordingly, there is no contradiction between results.

To better quantify the system performance, the receiver operating characteristics (ROC) curve is drawn for each pair in Fig. . The efficacy of each step of this experiment is depicted using True Positive Rate (TPR) and False Positive Rate (FPR) 10 . Thus, some measures are defined. True Positive (TP) is true command sent to the receiver and true movement detected; False negative (FN) is true command sent to the receiver but false movement detected; False Positive (FP) is false command sent to the receiver but true movement detected; True negative (TN) is false command sent to the receiver and false movement detected. It is noteworthy that the mentioned values have been calculated solely regardless of prior steps. The accuracy of the test depends on how well the test separates the group being tested into detected and not detected trials. Accuracy is measured by calculating the area under the ROC curve (AUC). The closeness of the curve to the top-left corner of the ROC chart depicts superior performance. This figure gives a good comparison between control session and experimental sessions, with step-wise contrast. The lowest efficacy in all stages of the experiment is 0.6 of full AUC corresponding to step 2 of pair 3, which is higher than the area of control experiment (0.5). Although the largest area under the curve is 1, the best MMI + BBI efficacy is related to pair 5 with overall 0.94 ± 0.08 of AUC. Step-wise analysis illustrates that step 3 with 1.0, step 4 with 0.977 ± 0.028, step 5 with 0.924 ± 0.044, and step 2 with 0.68 ± 0.09 of AUC present best to worst performances, respectively. Conclusively, test experiments for all pairs had a larger AUC than control ones, which is expected from a working system.

Considering the hierarchy of steps, efficiency of each step independently without the influence of previous one needs to be calculated separately. To address this, analysis of variance (ANOVA) of independency ratio over different steps using results from 20 sessions of the experiment is conducted. Stepwise ANOVA yields a p-value of 0.00025 which is dramatically smaller than the level of significance (alpha = 0.05). Thus, the null hypothesis, namely “the efficiency of each step is based on chance”, is rejected and results are statistically significant (F-value 5.908). The first step was excluded from IR assessment, since its independent accuracy is unconditionally pertinent to subject-related BCI performance, rather than system performance. Therefore, calculations started from second step, which illustrated IR = 0.87 ± 0.12 with 0.015 variance, while the third step perfectly completed without any misses (IR = 1). Meanwhile, fourth and fifth steps, demonstrated IR = 0.97 ± 0.06 with 0.003 variance and IR = 0.878 ± 0.13 with 0.01 variance, respectively. Hence, individual steps functioned efficiently.

The overall RA is 69.7 ± 8.55% and 87.37 ± 9.7% for BBI and MMI segments, respectively, which suggests muscle to muscle interface with the current mechanism is more reliable and accurate than the current BBI mechanism. The total response rate of the mechanism over 20 session shows 85.7 ± 12.13% accuracy. Notably, the accuracy of TMS-induced EMG detection is always 100%, owing to the fact that whenever right-hand motor imagery detection is true, a trigger is sent to TMS. The same situation is valid for EMG-triggered FES detections. The difference between pairs mostly depends on hand motor imagery and passive movement induced by FES classification accuracy. The rest of the system is properly operating which is in the prospect of a reliable BBI and MMI system. Overall RA of FES-induced BCI classification in the offline analysis have the average of 84.5% while for motor imagery classification is 63% totally (for stop and start commands). Conclusively, the overall FES-induced BCI classification accuracy was higher than motor imagery.

where the number of trials with a true response for a specific step is N TT and the number of related trials in each step is N RT . Here dependent RA corresponds to responses recorded for each step with respect to the prior step, for instance, the accuracy for BCI-triggered TMS detection is calculated with respect to the number of true right-hand motor imagery classification, therefore, we call aforementioned accuracies dependent accuracy. Meanwhile, independent RA is defined to show the efficiency of each individual step without any respect to its previous step. Since the number of visual cues on the right and left in each session is equal, the number of related trials for dependent BCI is 30 and for the rest of steps is 15. Dependent RA for each pair is shown in Fig. . As depicted from the figure, pair 5 has the best mean RA (80%) while the value of pair 3 is the worst (53%). Meanwhile, the highest RA reaches as high as 93%, occurred in the 3rd session of pair 5. Step-wise accuracies for each pair (Fig. ) are obtained, according to the results, MMI showed superior outcome to BBI for all pairs.

For right-hand true detection, the accuracy is calculated with the number of successful detections to the number of visual cues instructing the same hand motion imagination. Each session consists of 30 trials, among which half trials visual cue is on the right and half on the left. Figure illustrates overall independent response accuracy (RA) for 4 sessions (120 trials) of the experiment for each pair. Each loop is categorized into four independent steps and a dependent step of human-controlled paradigm. Accuracy for each step is calculated by:

The representative results are shown in Fig. . Brain activations are reflected by event-related de-synchronization (ERD) 9 , which is a normalized power attenuation in a specific frequency band (upper alpha rhythm 10–13(Hz)) with respect to a baseline time window (0.8–0.1(s) prior to the placement of cue). Motor cortical activations in form of 2-D head plots are illustrated to vividly describe the underlying mechanism of the brain and how it interacted with the mechanism. Head plots - drawn using FieldTrip, an open source MATLAB package - indicate activation spots over the brain according to electrode positioning. The majority of activations for both head plots are in the vicinity of C3 electrode over the contralateral hemispheric motor area of the right hand side. A time-frequency map of channel C3 is shown to illustrate a point to point brain activation in the contralateral hemisphere during MI task. Time-frequency plots are employed for this purpose which are plotted using EEGlab (open source MATLAB toolbox). Furthermore, sequential steps with their corresponding time delay are illustrated in Fig. . Accordingly, each loop takes 6.49 seconds in total to perform all steps (excluding rest time). Although BCI classification part had a 0.17 s delay, trigger circuitries did not introduce any detectable time delay while trigger signal transmission through the Internet showed 0.1 s latency. However, a large portion of the delay (4.5 s) corresponds to MI-BCI task and FES-evoked BCI. Despite the complexity of the mechanism design, hardware- (i.e. connections, network, circuits, processing times, etc.) and software-related (i.e. signal processing algorithm, device drivers, system software, etc.), latencies were negligible (merely 0.99 s). This fast response time corresponds to the short delay of BCI classification, EMG-triggered FES detection, and network data transfers which were approximately 0.17 s, 0.35 s, and 0.1 s, respectively. The interface between BBI and MMI was established using natural neural network connection from contralateral hemispheric motor area to index finger muscle with a trifling delay.

Discussion

A novel and distinguishing communication method based on BBI and MMI is demonstrated herein. Our current results indicate the possibility of functionally combining natural and artificial neural pathways to establish a novel communication path. It has two sections, first, from one person’s brain towards another person’s brain and muscle, second, from one person’s brain to their own hand. This is accomplished by merely using noninvasive technologies with short latency and reliably working mechanism. In other words, the goal was to establish a composite neural pathway using human natural and BCI-based artificial neural pathway to transmit commands from a sender’s brain to his muscle. To this end, cortical activations decoded from EEG and muscle movements detected from EMG signal were used as inputs, while FES and TMS were used as outputs of the artificial neural system. Experimental outcomes showed up to 85% of mean system RA achieved and totally 37.44 ± 2.5 bits of mutual information transferred. Furthermore, mean IR of steps 2 to 5 is 0.96 ± 0.08 and 0.99 s delay, explicitly reveals the system reliability and relatively fast response time of different steps of the mechanism. The proposed system may provide interesting entertainment between two individuals. To our knowledge, intending to make a move and use natural neural pathway possessed by another person to induce movement command in the first person’s muscle is not investigated in any previous research work. The novel paradigm introduced in this paper proposes a new human-to-human communication way which engages human brain and muscles. A promising capability of this mechanism is the interaction between two subjects from any distance, using the Internet as a part of the artificial neural pathway. In this way, simultaneous learning of two or more persons who are far apart is conceivable. Although analogous works focused on BBI or MMI alone have been accomplished, the full interaction between two subjects has never been addressed in any previous works in the field. Our results suggest that the proposed procedure of the mechanism is performing satisfactorily. Thus, the current design of combining artificial and neural pathways to unconventionally communicate two persons is adequately sufficient and can be subject to find operational applications.

A substantial objective of the BCI-based research is literally for rehabilitation. In addition to its entertainment application, currently proposed system can be used fundamentally for rehabilitation. Four well-known technologies used in this system, BCI, TMS, EMG, and FES are popular rehabilitation technologies. In former studies, the four employed technologies are widely used independently. Some researchers have attempted to combine two technologies among them for rehabilitation. For instance, BCI have been combined with FES12 and TMS13. Additionally, MI-based BCI has been proven to effectively improve motor recovery14. BCI can decode the motion intention and control related devices, in this case, the patients are in a rather active role in the rehabilitation process. It might potentially enhance the rehabilitation of stroke-related motor impairments by promoting neuroplasticity and altering motor cortical areas. Therefore, rehabilitation is likely to be remarkably improved15. Our proposed method initially combined all the four methods in a closed-loop system. The two subjects in side A and side B can exchange and experience all the four kinds of rehabilitation technologies. Majority of current rehabilitation procedures are performed solely on a single patient, but our research can engage two patients to realize co-rehabilitation as double players in a game. The co-rehabilitation could be designed as a competitive way, which means both patients in side A and side B are able to start or stop the rehabilitation process. They can compare “who win or who lose”. They can also exchange their roles in side A and side B to play again. They may have strong desire to win, so the rehabilitation would be performed more enthusiastically. In such a way, the co-rehabilitation may become more attractive and interesting for the patients. In this work, a combination of BBI and MMI is accomplished to illustrate the possibility and applicability of this kind of rehabilitation, which is based on the reciprocal human to human interaction channel.

According to results, FES-evoked BCI demonstrated higher cortical activity within the demanded frequency bands (i.e. alpha and beta rhythm) compared to MI-based BCI as indicated by ERD plots depicted in Fig. and FES-evoked BCI classification accuracy. This phenomenon might rely on the effect of FES-generated sensory feedback, which activates cerebellum and increases its interaction with cortex. Electrical stimulation of motor nerve fibers may generate both orthodromic and antidromic impulse. An impulse can cause depolarization of horn cells which leads to conductivity increase between pyramidal tract axons and anterior horn cells. Therefore, FES can induce changes in the segmental level even in people with lesioned limb and capable of activating anterior horn cell repeatedly which leads to enhancement in the corticospinal excitability compared to MI alone16.

The present work is a pilot study which has some limitations and there is a vast room for future development. Although noninvasive signal acquisition and stimulation methods employed in this paper are easy to implement and portable, their inherent shortcomings restrict further development of the mechanism. For instance, assessment of different natural neural pathway (i.e. lower limb motor nerves, etc.) effect is barely practicable. In addition, according to network based brain activations, for a perfect BBI all activated cortical spots in the brain of the subject in side A should be stimulated in the brain of subject in side B with corresponding latencies. However, it is impractical to stimulate different spots using a single TMS coil with low latency or using several TMS coils for neighboring spots. This is due to the size of the coil which prevents the reachability to certain regions that are spatially close to each other. Therefore, merely the physical response of the motor unit was used to determine which brain area to be excited. Moreover, only alpha and beta frequency bands are used to determine brain activations. This is mostly based on low signal to noise ratio of EEG signal which limits usability of features in other frequency bands (i.e. gamma band). Hence, noninvasive methods are replaceable with more accurate invasive technologies, i.e. Electrocorticography (ECoG), to accomplish a real task consisting of complex movements. Furthermore, implementation of the mechanism is technically complicated, this may pragmatically cause technical deficits while putting it into real-world application. Furthermore, simultaneous FES stimulation and MI suggested for maximizing short term neuroplastic effects17, while in our study they have been performed separately, which may be subject to modification in future rehabilitation practices.

Current software and hardware infrastructure is capable of being adopted for more than two healthy or disabled subjects to have interaction (e.g. social communication) with each other. Therefore, co-rehabilitation between two or more paralyzed patients simultaneously using the proposed mechanism can be further investigated. Besides, the response rate of the mechanism was directly interdependent to BCI performance. Thus, advanced sensorimotor pattern extraction algorithms to enhance the performance of the system and fundamental research on the exploration of the brain network that will provide more and effective control modes for BCI, are still in high demand. The current system design is one out of many possibilities. For instance, a possible substitution for BCI part is online ERD calculation. In this case, instead of using left or right hand EEG signal classification, attenuation or increase of ERD value in electrodes neighboring motor area can be used as BCI output and generate a trigger for TMS. Accordingly, TMS stimulation intensity could linearly be adapted with ERD value, meaning higher ERD value cause higher TMS stimulation intensity and vice versa. As a result, the amplitude of detected MEP will change linearly with TMS-induced movement, and accordingly, the FES stimulation may be modified and applies higher or lower stimulation intensity regarding MEP amplitude. Under this circumstance, both subjects will not only stop or start the mechanism but also control the range of movements which enhance the interaction level between two subjects.