Surface Electrodes (SEs) vs. Implantable Electrodes (IEs)
Currently, the most advanced commercial robotic prosthesesa
are limited to the proportional control (speed and strength) of basic movements (i.e. opening and closing the hand). They all use surface electrodes to read myoelectric signals (MES) from relatively strong muscle contractions, which is inefficient and unnatural for the patient, thereby complicating prosthesis acceptance [91
]. Several solutions have been implemented to cope with the poor myoelectric signals provided by surface electrodes; for instance, in order to control two different devices (e.g. hand and wrist), the fast contraction of all the muscles involved is used to switch from one device to the other. Although practical, an approach of this kind is still unnatural and cumbersome.
The lack of classification algorithms [11
] and control systems [92
] was once the main issue in accomplishing an advanced prosthetic control. Nowadays, several researchers have demonstrated that it is possible to identify finger and hand positioning using a variety of pattern recognition algorithms, such as artificial neural networks (ANNs), support vector machine (SVM), hidden Markov models (HMM), wavelets and so on [91
Now that pattern recognition algorithms and hardware for real-time control are available, the long-term stability of the biopotential signals has become the major issue. Signal stability is strongly related to the acquisition method and source. The following considerations must be kept in mind when designing an acquisition strategy for natural control.
1. The control source. The control origin has an impact on how natural it is for the patient to produce the information required for a given movement. In other words, the control source must be physiologically appropriate in order to be natural for the user.
2. The information content. This relates to the information required for the simultaneous control of different degrees of freedom. Simultaneous control is an essential feature when it comes to mimicking the natural physiological systems.
3. The long-term consistency of the signals.
Although the PRAs and control systems are designed to be robust and tolerate a certain amount of noise, they are still dependent on the consistency of the inputs [10
]. Numerous failures, such as low responsiveness and false predictions, tend to frustrate the user [99
The following is a non-exhaustive list of the issues found when implementing more advanced prosthetic control based on PRAs. They all conflict with at least one of the subsequent considerations, especially the long-term consistency of the signals.
· Signals recorded using SEs change dramatically with the environmental conditions, i.e. sweating.
· SEs cannot remain in place indefinitely due to skin-related issues and they therefore require daily placement.
· SEs cannot be placed consistently in the same location after removing the prosthesis. A different placement of the electrodes usually requires the re-training of the PRA, or recalibration.
· Artifacts are very easily generated when using SEs due to limb movement and electrode liftoff.
· Patients need to have a minimum level of myoelectric signals to become candidates for using myoelectric prostheses. This is not always the case, depending on the amputation level and the muscle surface left for the electrode placement.
· A wide limb surface area needs to be covered to have enough control signals.
· Control sites are insufficient in patients with high amputation levels. Moreover, as a rule, the few sites available are physiologically inappropriate.
· Muscle imbalance could be created if the electrodes are improperly placed, resulting in some muscles being more exercised than others. In the long term, this will cause the signals from the larger muscles to mask those from the smaller ones. Muscle imbalance can also cause prosthesis socket instability [11
· The lack of natural feedback, or any feedback apart from visual, complicates the use and acceptance of the prosthetic device.
It has been suggested that the long-term stability of bioelectric signals required for a precise control of several degrees of freedom, is the major issue in robotic prostheses [3
]. By reviewing the previous list, it is reasonable to argue that this can be mainly attributed to the use of SEs. As mentioned before, the currently available commercial prostheses, and most of the research on prosthetic control, use SEs [10
], mainly because they are easy to manufacture and non-invasive. On the other hand, the way in which the nature of SEs inherently impacts the problems mentioned above is obvious. This raises the question of how much better the IEs are in comparison to their surface counterparts. Selecting between SEs and IEs involves different trade-offs, such as the global sensitivity and specificity of the recorded bioelectric potentials. Table summarizes some of the most important differences.
Comparison between surface and implantable electrodes
It has been widely suggested that implantable electrodes acquire signals of better quality [17
], but, when it comes to improvements in pattern recognition, Hargrove et al.
did not find any significant difference between using 16 down to 4 SEs or 6 intramuscular electrodes while classifying 10 different types of isometric contraction [101
]. Farrell and Weir also conclude that the selection between IEs and SEs should be made according to clinical considerations. This was argued after performing a more in-depth study comparing surface and intramuscular electrodes. In this study, key muscles were targeted and compared with a symmetric electrode distribution [102
]. It is worth noting that the latter studies were performed using intramuscular electrodes and the question of whether using implanted nerve-based electrodes
will improve pattern classification is still unanswered. It is possible to argue that the information contained in the nerves that innervated the missing muscles could make a difference in classification performance. Furthermore, IEs, specifically nerve-based electrodes, would provide more physiologically appropriate information, and would also be suitable for a closed-loop control through stimulation of the afferent nerves (feedback).
Independently of the latter assumptions and considering that both types of electrode are able to perform the same control, IEs still solve most of the practical problems associated with SEs. It has been shown that by using implanted electrodes, the signals are more consistent over time and less affected by surrounding noise [103
]. Furthermore, it has been reported that patients using multifunctional prostheses with SEs experience fatigue after 5 to 30 min of continuous use [104
]. This can be attributed to the large muscle effort that is required to produce suitable myoelectric signals. This effort is considerably less when using IEs, as there is no skin and fat between the electrode and the source. Fatigue is not only unpleasant for the patient, it also complicates pattern recognition.
Several authors have found that at least 4 SEs are required to reach a good degree of classification accuracy when using PRAs [101
]. In the case of intramuscular electrodes, it was suggested that more than 4 electrodes are necessary, since the recording is more localized [102
]. It is noteworthy that intramuscular electrodes are more selective than extramuscular ones (such as epimysial) and for this reason, the question of whether more extramuscular electrodes are required for the same purpose still needs to be investigated.
The latter exemplifies the tradeoff between the global sensitivity and specificity of the recorded signals. In a direct control scheme where one signal is paired to one action, a localized measurement is more desirable, and provided that there are as many control sites as movements to control, this would be the most straightforward path to achieve simultaneous control.
Simultaneous control of 2 degrees of freedom using the direct control scheme was first demonstrated by Kuiken et al.
]. This was only possible due to the Targeted Muscle Reinnervation (TMR) procedure which resulted in new and independent control sites. Unfortunately, even in TMR patients with augmented myoelectric control possibilities, it is not always possible to isolate MES satisfactorily from surface recordings. This is thus in favor of the development of implantable MBEs.
Although not as extensive as in sequential control, PRAs and biologically inspired algorithms have been used to predict simultaneous limb movements since 1973 by Herberts et al.
], and more recently by Jiang et al.
], and Muceli et al.
]. Again, these were laboratory experiments using surface electrodes that could potentially be translate into long-term implementations using implantable electrodes.
Kilgore et al.
have shown that the durability of implantable electrodes, leads and connectors is no longer a major concern in implantable neuroprostheses. They looked at 238 electrodes implanted from 3 to 16 years, where only 3 reported failure. Furthermore, most of the leads crossed different joints, which increased the level of mechanical stress, and yet a 98.7% probability of being functional after 16 years was still reached, even in contracting skeletal muscles [22
]. Although these were used in neurostimulators, the safety, biocompatibility and recording performance can be seamlessly applied to prosthetic control. Prosthetic control, however, faces the issue of permanent communication with an outer-body device. Neurostimulators, on the other hand, can be completely implanted and rarely require communication with the outside.
Muscle-based Electrodes (MBEs) vs. Nerve-based Electrodes (NBEs)
In the case of an amputation, most of the muscles required for dexterity are lost. However, the axons that used to innervate those muscles remain indefinitely viable on the nerves close to the stump [110
]. It has been suggested that nerve signals could be used to control prostheses when it was found that the modulation of severed efferent neurons was still related to identified movements [111
]. A feasibility study from 1982 conducted in the median, ulnar and radial nerve of a below-elbow amputee supported the latter theory [112
]. A more recent study by Jia et al.
showed that, even after 28 months of being amputated, a patient was able to generate neural activity related to the phantom limb movement [113
]. These signals were enough to control simple movements in a robotic prosthesis, although the patient only trained for 2 weeks prior to the experiment. Kuiken et al.
’s work on TMR provides another strong argument in favor of the viability of nerves after amputation [114
Currently, as many as 3 different neuroelectric signals, both efferent and afferent, have been differentiated using pattern recognition algorithms. Micera et al.
classified 3 movements to control a robotic hand using 4 intrafascicular electrodes (tf-LIFE4) [71
] and Raspopovic et al.
showed that the identification of 3 afferent stimuli is possible using single-channel cuff electrodes [43
Theoretically, NBEs are very attractive because of the large amount of information they would make it possible to retrieve. Signals to several muscles can be obtained from a single nerve and several MBEs would therefore be necessary to produce the same information as one NBE with several contacts. It is noteworthy that the latter would not apply if the electrodes were placed in single fascicles. In both cases, NBEs could still provide information from missing muscles, information that could even be recorded from individual fascicles [65
]. Furthermore, it has been observed that neuroelectric signals (NES) show greater shape regularity than the MES [36
]. This facilitates pattern recognition and thereby improves the stability of the control system.
Since the MES are in the order of mV and the NES are in the order of μV
, the former are easier to record. Another complication for NBEs is that the MES of surrounding muscles can mask the NES, further complicating their recording. On the other hand, biosignals peak in different spectra (MES 200 Hz, NES 2,000 Hz), making filtering feasible [17
]. In some cases, a band-pass filter from 1,000 Hz to 10,000 Hz or a high-roll-off high-pass filter have been sufficient to cut off the MES [31
The variable position of the MBEs in relation to the muscle contraction results in a position-dependent signal [18
]. The latter is not an issue in NBEs, which are also more efficient for stimulation purposes because they use less current and have better selectivity than the MBEs [116
]. Furthermore, NBEs eliminate the problems of mechanical stress that plague MBEs, making NBEs more stable over time and giving both the electrode and the leads a longer life expectancy [28
Finally, NBEs are inherently more suitable for providing feedback that is perceived in a more natural way by the patient. Closed-loop control using different NBEs has been reported in animal models [115
] and human patients [40
Suitable IEs for long-term and natural prosthetic control
-type implantable electrodes (i.e. intraneural) break and damage the tissue where they are implanted and therefore have a higher associated risk than the extra
]. The need for devices such as the ACTIN resulted from the degradation of the long-term recording qualities of intrafascicular electrodes as a consequence of the fibrous encapsulation that takes place [66
]. While it has been suggested that encapsulation is beneficial in epimysial and cuff electrodes [18
], it is a major issue in intraneural electrodes [72
]. This is mainly due the higher impact of attenuation on the smaller signals. Nevertheless, most of the published research on electrodes with direct application to robotic prostheses has used this more invasive designs [26
]. On the clinical side, however, extra
-type electrodes such as the epimysial and cuff, have proved their safety and functionality for several years in different applications [19
]. It is worth noting that a screened epimysial electrode design as in Figure ??, with a short inter-electrode distance, would result in a very localized sensitivity. This would practically be equivalent to the use an intramuscular electrode for direct control purposes. Furthermore, low spatial resolution (high selectivity), also implies low information content. Selectivity and global sensitivity both highly depend on the electrode design. Therefore, electrodes must be carefully selected considering the available control sites and the desired control strategy.
It has been shown that neural interfaces can provide enough information to allow PRAs to identify different movements [43
]. It seems fair to argue that the combination of implanted NBEs and MBEs has the potential to solve the 3 main problems mentioned in the surface vs implantable electrodes discussion.
1. The control source. When using muscle- and nerve-based interfaces, the origin of the control signals is the same as that of an intact limb. Especially with neural interfaces where the information would be physiologically more appropriate even in high-level amputees. Feedback would be also transmitted directly in the remaining natural conductors.
2. The information content. Nerves contain all the information required for natural control. Arguments in favor of their viability after amputation have been given in this work, as well as citations of different experiments in which recordings and stimulations had been achieved. MBEs could well complement the control system increasing the number of sites for control. In some cases such as in low trans-radial amputees, MBEs alone provide a considerable amount of control sites.
3. The long-term consistency of the signals. Implantable electrodes, especially the epimysial and cuff electrodes, have been shown to be stable in long-term implementations where the consistency of the electrode impedance has been observed. This feature directly relates to the stability of bio-electric signals.
In cases in which suitable nerve recordings for pattern recognition cannot be obtained due to technical limitations, it should still be possible to use epimysial electrodes to achieve the stability required for long-term implementation. Despite the fact that information from muscles might not be sufficient to achieve dexterity, since most of the required muscles are lost in an amputee, an approach of this kind will still dramatically increase the functionality of the prosthesis by increasing the number of movements and allowing simultaneous control. A simple control scheme pairing single movements with individual, but well isolated myoelectric signals, would be enough to produce simultaneous movements. Finally, the combination of nerve and muscle recordings could ultimately be used to achieve long-term stable and natural prosthetic control.
It is worth mentioning that electrode technology continues to progress and new materials, coatings or surface treatments such as PEDOT [88
], silk fibroin [90
] and ion-selective membranes [89
] could eventually replace or be applied to standard epimysial and cuff electrodes. These are just a few examples of potential future improvements that could eventually enter the realm of clinical applications.