Therapeutic robotics started about 20 years ago. Because robots can be used to reproduce different multi-sensory interactive scenarios, applying robot-assisted therapy allows us to customize the interventions on individual physical impairments [9, 30, 42, 43], and to also provide precise control over a large number of physical variables--haptic, visual, and auditory cues--that influence motor behavior [10, 32]. Here we explored the efficacy of using EMG signals, not as a measure of the strength of specific groups of muscles [12, 42] but as a way to detect intentions to move toward a certain direction. This classification could be used in future applications to enable the control of assistive and rehabilitation robotic devices. With this aim in mind, we examined whether we could build a static functional map of the EMG activation patterns for point-to-point reaching movements located on the horizontal plane environment. We used the SVM algorithm to predict the intended motion direction with the highest possible accuracy and to understand whether and how the inter- and intra-subjects’ variability could affect repeatability of the measurements.
Limits of EMG pattern recognition
The approach was quite successful for healthy subjects. In this case, we were able to achieve a classification accuracy of 89.6 ± 4.4% (TEST 2, Table 2). Even more interesting, we employed the data from all subjects and did not train the algorithm on individual data. Inter- and intra-subject’s variability was not a critical factor affecting the classifier performance, as shown by the higher classification rates obtained with TESTS 2, 3 and 4 reported in Table 2. Accuracy increased significantly, i.e., up to almost 97.5% of the success rate when the algorithm was trained and tested on individual data or when validating the SVM with our best subjects. However, we were not able to obtain 100% correct predictions in any case. While the accuracy could be increased further (for instance with the use of more sophisticated techniques [19, 36, 44]), the results obtained in the present analysis are in line with different techniques used in the past for similar applications. For example, it has been reported an accuracy close to 100% using a neuro-fuzzy classifier to detect planar arm movements similar to those tested in the present study . In the case of EMG based control of prosthetic devices, several studies reported a success rate ranging between 94-99% [17, 45, 46]. Other classification algorithms, such as the Hidden Markov Models (HHM) and Higher Order Statistics (HSO), have also been tested but the results have not been very promising so far [47, 48]. Finally, brain computer interfaces (BCI) used to command computer devices with upper extremity muscle activity achieved a remarkable 96-97% recognition of individual intentions . Of note, in all the cited studies the classifier was tuned to the specific individual, as for TEST 1 in the present analysis.
Overall the present analysis showed that in the case of healthy subjects, beside the presence of episodic abnormal co-activational patterns at the base of SVM misclassification, EMGs can provide a reliable map of the coactivation relationships between groups of muscles.
EMG pattern recognition in robot-mediated neuro-rehabilitation
In the case of stroke patients, the classifier accuracy dropped dramatically and ranged between 25% and 45% (Table 5). Mainly, the classifier failed to recognize motion toward the East and the West directions (i.e., the average accuracy rates were 8.6% in both cases). Consistent differences in the EMG patterns between the two groups were also observed (see Figures 2, 3, and 4). The analysis of the CoE directional distribution revealed the presence of individual abnormal patterns of activation--spastic muscular restraint, muscle synergies, lack of elbow extension, weakness of specific muscle coordination deficits--in accordance with previous studies [8, 12, 26, 27, 33, 41, 48–52]. For instance, the simultaneous activation of the shoulder abduction and elbow flexion, i.e., flexor synergy [41, 53, 54], was present in several subjects (Patient #1, #2, #3, #4 and #5). Both DM and DP muscles were no longer activated with the elbow extensor and the TR muscle, and showed a high CoE coefficient when attempting to move toward the South direction. This could explain the misclassification rates observed when attempting to move toward the North (i.e., 68.6%) and East directions (91.4%), as well as the large amount of trials misclassified in the South direction (Table 5).
The number of correct classifications increased when the system was trained and tested on individual data (Tables 6 and 7). The best classifier’s performances were achieved when a larger number of muscles was used, i.e., including the LT and TM muscles (Table 7), in accordance with previous studies which reported an increase in the involvement of the trunk in post stroke upper limb movement coordination . With this approach it was possible to achieve up to 83.3% of correct classification, as in the case of Patient #6. Nevertheless, for the other patients the accuracy was not as high as in the case of GROUP I and ranged between 36.7 -70%. In this context, it is important to note that an accuracy of 36.7% was close to the probability of getting a particular direction by chance alone, which in the present case was 25% given that the classifier had to discern between four very distinct possible classes. Overall the accuracy in the East and West directions increased up to 59.8% and 64.9%, respectively. To achieve a better classification performance, SVM generated the hyper planes that increase the separation between the different classes. The solutions exploited varied according to the subject specific EMG signal characteristics, that is, the distribution of the HIST vectors in the task space. One of the possible drawbacks of this strategy was that the system might have recognized the intention to move toward a certain direction from the pathological incorrect schemes or the stereotypical coupling muscle patterns. Moreover, the method was affected by the presence of large variability of the EMG features within each class. In fact, each muscle presented activation over a broader range of directions compared to healthy subject groups as shown by the large confident intervals of the CoE parameter distribution in Figure 2. It has been largely documented that stroke patients present an increased movement variability with respect to normal population [49, 53, 55]. The end point kinematic is often characterized by large path errors and a speed profile composed of several low amplitude peaks, markers of discrete sub-movements that underlie deficits in motor control [33, 41, 54–57]. We also observed these characteristics in the performance of the stroke patients enrolled in this study (Figure 4). In this context, compensatory strategies and corrective actions exploited by patients in their attempt to complete the task might have been changed from trial to trial as also reported in previous studies .
Summing-up, the approach was not successful in the case of pathological subjects. The aimed direction was not predictable with a sufficient accuracy whether using the general map extracted from data of normal subjects (i.e., reference model), or tuning the classifier on each individual.
Some limitations of the study should be highlighted at this point. For instance, we only recorded a small subset of all the possible trunk and upper limb muscles. In the case of GROUP I, this approach worked quite well. However, in the case of GROUP II the number of trunk and back muscles should have been increased to account for the presence of compensatory strategies in the classifications extracted with the SVM. Proper electrodes placement was a critical issue when dealing with patients due to the weak EMG signals. Moreover, patients were often overweight due to physical inactivity with a subsequent EMG signal decay, which further affected classification accuracy. Additionally, we could not rule out the presence of fatigue at the base of intra-subject variability in the performance of stroke survivors. Finally, in the present study we used a statistical classifier. Neural networks might also have been employed as an alternative, given the good performances in the pattern recognition described in literature [19, 34]. A recent study  however, reported no significant differences in the classification between the two approaches.
New therapies and solutions
The results of this study are in line with a previous work from Lee and colleagues , which applied a subject-specific EMG patterns classification technique to discern the intent of stroke patient in performing six different manual tasks. To this aim they used the LDA (Linear Discriminant Analysis) classifier on segments of data of 150 ms in duration shifted in 100 ms increments. While a similar approach could be applied also to the present analysis to improve recognition accuracy, classification performances were nonetheless comparable to those herein reported. Specifically, mean recognition accuracy was 71.3% for moderately impaired subjects and 37.9% for severely impaired stroke. The poor classifier performances observed in the case of the severely impaired patient group, posits an important issue in our view. In fact, there might be a problem when using the classification to enable volitional control of assistive devices. For instance, if the patient produces uncoordinated activation patterns, which are not correctly interpreted by the classifier system, the robot could move in an undesired way. Moreover, having the system calibrated on the data from each patient might be not practical in clinical application. It took us over 3 hours to collect a patient’s data with the support of a research engineer and a research therapist. While it is possible to automate the process to collect the data by a regular clinician within an hour, it will require significant design effort and, in lieu of our poor results classifying 4 very distinct classes in severe stroke, lead to a low cost/benefit. Finally, there is the theoretical possibility of “bad” plasticity: the classifier could recognize the intended motion direction while reinforcing some unwanted pathological incorrect schemes [58, 59].
It is often speculated that the process of motor recovery either involves spared tissue on the motor and sensory areas in the lesioned hemisphere or it enhances activity in pre-existing motor networks in the unaffected hemisphere. Several solutions employing EMG are possible. For example, it could be possible to calibrate the system on patient-specific characteristics and to have the clinician select the admissible co-activation pattern that represents an effective strategy to inhibit pathological scheme. The control system of the robot could then be implemented in order to train patients to always reproduce the same correct pattern. The rationale underlying this approach is that learning may be elicited providing the subject a feedback on the muscle forces he has to produce to achieve the desired trajectory.