Low-Cost Textile Myoelectric Control of Knee-Ankle-Foot-Orthosis

The poorest populations in the world have the highest prevalence of lower limb disabilities, and lack of access to healthcare prevents many from lifting themselves out of poverty. This is particularly true for the large population of poliomyelitis-affected inhabitants of India, whose quality of life would benefit substantially from the provision of affordable, yet modern, dynamic knee-ankle-foot orthoses to assist in ambulation. To this end, this paper reports a study into the use of a low-cost, textile-based sensor interface for the myoelectric control of lower limb orthoses in restoring gait function. It reports experiments examining the accuracy with which gait events in the healthy limb (e.g., heel strike, toe-off) can be detected through the textile interface, with a view to triggering discrete control modes of a smart orthosis (i.e., knee lock and release) to support the atrophied limb during walking. Results show that prediction accuracy through the proposed interface (∼ 70%) approaches that of more traditional medical-grade sensors, despite its substantially lower cost.


I. INTRODUCTION
One billion people globally have a disability, and 80% of them live in developing countries [1]. People with disabilities (PwDs) are over-represented amongst the persistently poor, and are less likely than others to be able to move themselves out of poverty [2]. In India specifically, around 19 million people are estimated to have disabilities, of which around 10.3 million people have a locomotive disability [3].
Knee-ankle-foot orthoses (KAFOs) are commonly used to assist patients with gait dysfunction and allow them to recover mobility and independence. Not all people with a lower limb disability are prescribed a KAFO, and there is lack of data showing exact number of users in India. However, the International Committee of the Red Cross estimates that there are more than 10 million people in India in need of a KAFO [4]. Polio survivors, especially, are in need because their lower limb muscles are usually too weak to be able to support their body weight.
In less economically developed countries, passive-KAFOs-essentially, rudimentary leg-braces that allow the knee joint to be manually locked-are the most commonly used. However, these devices are the bare minimum for enabling locomotion, and healthy gait is not restored. In developed countries, electromechanical KAFOs, or stancecontrol-KAFOs, allow for knee flexion during the swing phase, along with stance phase knee stability. These devices can often be sensorised, to provide more refined control, 1   enabling patients to, for example, descend slopes and stairs and walk on uneven terrain [5]. However, the high cost of such devices limits their accessibility.
A key cost in the creation of electromechanical KAFOs is their sensorisation for purposes of automatic control. Typically, such devices use either inertial measurement units (IMUs) [6] or surface electromyography (SEMG) [7] to detect gait events such as heel-strike and toe-off, in order to determine the appropriate control mode of the KAFO. SEMG is particularly promising as a low-cost sensing modality, since it has been shown to be effective as a means to control KAFOs by decoding the muscle activity of the healthy leg [8], however, for widespread deployment the purchase and maintenance costs must be reduced.
Currently, several SEMG electrodes are commercially available at different price points (see Table I). Nongelled, reusable Ag/AgCl are available at approximately £11 per unit. Disposable, adhesive, pre-gelled electrodes (e.g., Ag/AgCl Covidien Kendall disposable electrodes [9]) are also available, at the much cheaper price of approximately £0.26 per unit. However, the use of disposable electrodes requires users to have a stable supply of them for daily reattachment to the limb, something that is often costly and difficult to maintain in developing economies.
In contrast, recent research into smart textiles has resulted in a more affordable alternative to these traditional SEMG systems. Specifically, the use of conductive yarns embroidered onto a fabric substrate has been seen to be effective in creating low-cost, flexible and reusable electrodes, suitable for SEMG [10]. These systems are considerably cheaper than conventional systems: the cost in raw materials of a single electrode is around £0.16, and the fact that the electrodes are reusable drives the cost even lower. However, while they have been shown to be effective in applications involving affordable upper-limb prostheses [11], their use in orthotics is so far untested.
In this study, the use of embroidered textile SEMG sensors is investigated with a view to assessing their suitability as the myoelectric interface of a KAFO for control of the knee joint (see Fig. 1). Experiments are reported for N = 3 healthy subjects, in which the performance of (i) textile-based, and (ii) conventional gel-based SEMG systems are assessed for decoding patterns of muscle activation corresponding to salient events in the gait cycle (heel-strike and toe-off) during normal walking. Secondly, the possibility to predict gait events of one leg with SEMG data collected from the other leg is explored. Statistical classification is applied to perform the detection of heel-strike and toe-off and the decoding accuracy measured. The results indicate that, in this population, there is no significant difference in performance between the different electrode types with overall approximately 12% difference of accuracy in gait event detection, suggesting the feasibility of textile sensors as an alternative control interface for affordable smart KAFOs.

A. KAFOs for Polio patients
KAFOs are long-term assistive devices to augment the functionality of multiple lower-limb segments. The reasons for prescribing KAFOs are heterogeneous, and they are commonly used to assist abnormal walking gaits, specifically abnormalities in the control of the knee and ankle joints. The etiologies of conditions that require a KAFO can include traumatic injury, neural damage such as spinal cord injury or multiple sclerosis, or muscle weakening diseases such as poliomyelitis (polio) [12].
Polio survivors in particular often require the use of orthotic devices to counteract muscle weakness. These are the key assistive technology for improving quality of life because they allow users to live independently [13], and increase the economic well-being of themselves and their dependants [14]. While polio vaccinations have reduced the number of new cases to less than 100 globally in 2018 [15], the life-long effects on the approximately 15-20 million global polio survivors [16] remains prevalent. Additionally, symptoms do not remain stable: a revival of symptoms known as post-polio syndrome can affect up to 20-30% of survivors [17] 15-40 years after initial contraction [18].
Research and development into assistive technologies for maintaining a person's quality of life, have resulted in a vast variety of devices to counter gait abnormalities. Generally, KAFOs can be classified into three main groups [19]: (i) Passive: where an orthotic knee joint remains locked during ambulation, and unlocked manually for sitting. (ii) Stancecontrol: where the knee joint is locked only during the weight bearing phase of gait via a passive-mechanism and knee flexion happens during the swing phase. (iii) Dynamic: where actuation mechanisms such as springs or motors are used to control joint motion in response to sensed stages in the gait cycle. Common commercial devices include the FreeWalk from Ottobock [20], the Stride from Becker Orthopedic [21] and the C-Brace from Ottobock [22]. These devices use joint kinematics and kinetics information to provide support in stance phase and to allow free knee motion in swing. Such control presents a big improvement over a permanently locked joint that disrupts the fluidity and biomechanics of normal walking [5].
However, while these technologies are readily available in developed countries, countries which have the largest populations with polio afflictions (e.g., Nigeria, Pakistan, India [15], [23]), are also amongst those with the largest socioeconomic inequalities in terms of access to healthcare. Additionally, these countries also have the youngest populations of polio survivors [24] due to global vaccination programs in 1988 to eradicate the virus. The fact that the demographic has a large young percentage is worrying, as not only is the economic power for the next generation limited, but as this demographic reaches middle and old age, an increase in individuals with post-polio syndrome will place additional demands on healthcare institutions [16]. As such, the demand for orthotic devices may increase in coming years, both in terms of first devices for patients whose condition was not previously severe enough to require one, and for patients who require more complex orthotics to handle changes in the pathology.

B. Low-cost EMG controlled KAFOs
Low-cost orthotic devices are required to enable the poorest population with lower limb disabilities to perform everyday ADLs. As such, passive-KAFOs are most commonly used, as these are the least expensive due to their simple design. However, users often experience difficulty in using the KAFO reliably in everyday situations, for example, operating the mechanism in a crowded location [13]. To solve these issues, dynamic KAFOs allow for automatic locking and unlocking of the knee joint. The SEMG-driven solutions in particular are well-suited to low-cost human motion analysis systems because they do not require highend electrical engineering in their construction.
The use of SEMG on the healthy leg to control an assistive device that supports the other, affected leg has been explored for simple motion such as squats [8]. However, nobody has explored the possibility of controlling an orthosis supporting an affected leg with SEMG data from its healthy counterpart during gait. With prior knowledge that normal gait (i.e., walking in a straight direction) is cyclic with a specific muscle activity pattern [25], there is a potential of  decoding SEMG signals from a healthy leg to control the knee joint of a passive-KAFO affixed to the affected leg in order to allow for knee flexion during the swing phase and locked knee during the stance phase using low-cost sensors. This solution is of interest especially for polio patients, who cannot use the muscles of their disabled leg.
Recently, advances in the creation of embroidered electrodes [10] have resulted in a low-cost SEMG sensing system, which can be used to perform motion classification in healthy subjects [26] or predict gestures of the phantomlimb in amputees [11]. These electrodes can be made from inexpensive conductive textiles, and provide many advantages over standard gel or metal-plate electrodes, such as re-usability and the possibility of local manufacture. Also, as they can be directly sewn into clothing at the locations necessary for measuring the muscles targeted, the need for a professional who is trained in sensor placement is reduced. These factors suggest that the use of these new sensors could be a promising route to driving down costs and thereby increasing access to smart KAFOs in the developing world with textile SEMG. However, their applicability in gait event prediction has not previously been researched, and there is no evidence to date that they will result in similar performance compared to standard gel-based electrodes.

III. MATERIALS AND METHODS
This paper reports an investigation on the feasibility of using surface electromyography data collected on the lowerlimb with embroidered electrodes for the prediction of two main events of gait through classification of muscle patterns: heel-strike and toe-off. 1

A. Experimental Setup
The experimental setup and protocols for collecting data are as follows.
1) Motion capture system: The motion capture system is a Phasespace Impulse X2. 8 cameras and a total of 13 active LED markers are used [27]. The data is collected at the rate of 120 frames per seconds. Any gaps in motion capture data due to LED obstruction during recording are filled using the interpolation algorithm in [28] because it has been shown to have good performance in data similar to that considered here (i.e., straight line walk with 6 markers per leg).
2) Textile electromyography acquisition system:  3.0 programmable sewing machine following the design guidelines presented in [29]. This electrode design is selected because it has shown an optimal trade-off between the electrical properties and the manufacturability of the produced electrodes. Once the CAD design file is converted to an embroidery file using the 6D embroidery software provided with the sewing machine, the electrode is sewn onto a fabric substrate (cotton linen and stabiliser) using stainless-steel conductive threads (Sparkfun DEV-11791, 3.28 Ωm − 1). An ordinary haberdashers snap fastener (Hemline H420.13.G, 13mm, gold brass rust-proof fastener) is sewn on the top side of the electrode to make the connection to a data acquisition device. An example electrode is shown in Fig. 2. For future signal processing (i.e., data segmentation), an additional pressure sensor (FlexiForce (FSR) Sensor [30]) is attached to the SEMG data acquisition device in order to detect when the heel of the leg where electromyography data is collected, hits the floor. Five channels of the BiTalino (r)evolution Plugged Kit BT are used to sample data from SEMG and pressure sensor synchronously at a of 1 kHz via Bluetooth (four channels for SEMG and one channel for pressure sensor recording). During data acquisition, the experimenter monitors the timedependent SEMG signal in real-time to verify good contact with the skin (poor contact is indicated by high-amplitude noise) using the OpenSignal software package (v.2017 2 ). The sensors used in the experiment are placed at desired position on the skin using kinesiology tape and velcro 3 .
3) Participants and Protocol: SEMG data and kinematic data are recorded from N = 3 healthy male subjects of ages 25, 24 and 28 years old. 4 The electromyography acquisition system is used to record the muscle activity from the right leg muscles and the motion capture system is used to record kinematic data from both legs. Firstly, 6 LED markers are positioned on each leg and one last marker is positioned on the sacrum according to the guidelines presented in [31]. The LEDs are placed on the femur (greater trochanter, and lateral epicondyle), knee (head of fibula), ankle (lateral malleolus), heel (the posterior surface of calcaneus) and toe (the head of the fifth metatarsal) of both legs. One marker is placed on the sacrum for data processing purposes [32].
The electrode placement is selected in accordance to the SENIAM recommendations [33]. Four pairs of electrodes are placed on the quadriceps, hamstrings, triceps surae and tibialis anterior. Previous work has shown this set of locations to give good gait phase classification [34]. The setup is shown in Fig. 3.
A trial consists of overground walking in a straight direction within the range of the motion capture system cameras (a distance of 4 metres), while data is recorded through the motion capture and SEMG acquisition devices. Participants are asked to walk at their self-selected walking speed, starting by a first step with the right leg. Before beginning the experiments, a mock trial is performed to make sure that the electrodes have a good contact with the skin and the LED markers are detected by the cameras. For each participant, 5 trials are performed for each electrode type.

B. Data Post-processing and Classification
To evaluate the potential of the textile electromyography system as a smart KAFO control interface, the data need to be decoded to predict key events of gait. To this end, here, the aim is to discriminate heel-strike (HS) and toe-off (TO), from the rest of gait (RG). This involves (i) the segmentation of the SEMG data, (ii) the computation of segment-wise features, and (iii) the classification of the gait events through a pattern recognition algorithm. The following describes these postprocessing steps in detail.
1) Signal segmentation: The first post-processing step consists of segmenting the raw SEMG signal preceding the events of HS and TO. Then the remaining data (corresponding to RG) is taken aside to be counted as a third class in the classification process.
In this study, two prediction scenarios are considered. In the first phase of this study, SEMG data taken from the right leg is used to predict gait events in that same leg, as a baseline measure of performance. In the second phase, SEMG data of the right leg is used to detect gait events of the left leg. The SEMG data is therefore segmented according to the leg in which gait events are detected. For segmentation, the SEMG and motion capture dataset are synchronised with the HS events. The HS events are detected in the SEMG dataset using the pressure sensor input, and in the motion capture dataset using (1). In the former case (same-leg detection), HS in the right leg is determined according to the pressure sensor by checking for the moment when its signal rises above zero. Using this as a reference, the motion capture data is used to find TO of the right leg, and HS and TO of the left leg, in the SEMG data. Specifically, the motion capture data is used to find (i) ∆ rHST O , (ii) ∆ lHST O , and (iii) ∆ rHSlHS , where ∆ rHST O is the time difference (in seconds) between HS and the following TO of the right leg, ∆ lHST O is the time difference between HS and the following TO of the left leg, and ∆ rHSlHS is the time difference between HS in the right leg and the previous HS of the left leg, see Fig. 4. These are computed from the motion capture data using the equations presented in [32] which represents the time at which there is a maximal displacement of the heel from the sacrum marker and a minimal displacement of the toe from the sacrum marker. After detection of the event of HS and TO in the motion capture data, the previously described ∆ rHST O , ∆ lHST O and ∆ rHSlHS can be computed as a simple time difference. After that, the timing of the HS and TO events for the two legs are computed as where T RHS is the time HS occurs for the right leg, detected with the pressure sensor, T LHS is the time of HS in the left leg, and T RT O , and T LT O is the time TO of right and left legs, respectively. Note that, the first and last HS and TO events are discarded for classification since they represents the initiation and end of the gait and are not in the framework of normal gait event detection [32].
Once the events are detected in the SEMG data, a segment of the length of 200ms is taken before the events of HS and TO detected in each trial. The rest of the data, RG, is then extracted and partitioned into segments of the length of 200ms. The whole process is summarised in Fig. 4.
2) Feature Selection and Classification: After signal segmentation, the event prediction is computed using statistical pattern recognition based on features extracted from the data segments. For this, a one-versus-one multi-class Support Vector Machine (SVM) with a Gaussian kernel is used [35].
In general, the data features given to the classifier affect the performance and speed of classification. In this study, following [34], the waveform length and standard deviation of each segment is used, having previously shown good accuracy.
To train the SVM, first, a 10-fold cross-validation is performed to find the optimal hyper-parameters (box-constraint and kernel scale) with respect to the accuracy where |T + | (respectively, |T − |) is the number of true positives (negatives) and |F + | (respectively, |F − |) is the number of false positives (negatives).
After the hyper-parameters are set, the classifier is trained on a random 70% of the feature vectors, with the remainder held back for testing. This procedure is repeated ten times for each data set (i.e., for each electrode type and for each subject) and the gait event recognition accuracy is computed each time according to (6).

IV. RESULTS
A. Gait Event Recognition with Embroidered Electrodes 1) Right leg event prediction: The classification results for the embroidered and gel-based electrodes for the right leg event prediction are summarised in Table II. The detection of HS and TO using the textile electrodes reaches a maximum accuracy of 86.78% for S4. In comparison to the textile electrodes, conventional gel electrodes achieve a higher accuracy (Table II, right column) with S4 and S5, with the difference being most noticeable in S5. This difference might be due to the fact that textile electrodes have not been able to collect as much information as the gel electrodes because this participant had very hairy legs which may not have allowed the electrode to lay completely on his skin. Overall, event detection was better with S4 probably because this participant had more developed muscle than the others. In contrast to the results for S3 and S5, better accuracy is seen with textile electrodes than with gel electrode for S3. Overall, the results suggest that textile electrodes can be used for event prediction in gait with fairly similar performances to gel-electrodes.
2) Contralateral leg event prediction: The classification results for predicting events in the left (contralateral) leg are summarised in Table III. There, it can be seen that the highest gait event classification accuracy using embroidered electrodes is 76.55%. Overall, in this experiment, the textile electrodes have a lower performance than the gel ones. The worst results for both electrode types are shown in S3, possibly due to low development of his leg muscles.The best results are seen for S4 and S5 with gel electrodes, however, it can also be said that the majority of predictions using the embroidered electrodes are accurate. Overall, the results show that contralateral leg event prediction is possible with embroidered electrodes although the performance is slightly lower than when using gel ones.

B. Classifier Evaluation
To investigate why the classifier performance is lower for the event prediction in the contralateral leg event with textile electrodes, the within-class accuracy (i.e., the performance of the classifier in distinguishing between HS, TO and RG) is presented in Fig. 5 for the participant S4. 5 It can be seen that the error largely comes from the fact that HS (and to a lesser extent, TO) events are predicted as RG. This means that the information taken prior to HS is insufficiently different to that used to train the RG class. A possible reason for this, is that the number of examples used to train the classifier is lower for HS (and TO) than RG. It therefore may be expected that the errors could be decreased by providing a better balance of positive and negative examples in the data set (see Table IV).

V. DISCUSSION
In this paper, the application of embroidered textile electrodes in detecting heel-strike and toe-off events during normal gait with a view for future control of the knee joint of a KAFOs is presented.
A difficulty in myoelectric control is the user reliance on disposable, gel-based electrodes for controlling the orthosis. This paper proposes the use of embroidered electrodes for muscle activity monitoring as a low-cost alternative, with the benefit of reusability and local manufacturing. To evaluate the proposed approach, SEMG data collected on the right leg is used to predict the HS and TO event of the same leg, and then the same data is used to predict the HS and TO event of the left (i.e., contralateral) leg. The results show that there is a potential to predict gait events either with gel or textile electrodes, albeit with comparatively lower performance in the latter.
To consolidate the results of this study an extension of work is required with a broader demography. However, it is anticipated that country-specific factors may make textile electrodes a more appealing choice. For example, the daily use of gel electrodes in the living condition experienced by the target populations of this study (e.g., hot weather in 5 The diagonal of the confusion matrix indicates the positive predictive rates (i.e., r + = |T + |/(|T + | + |F + |)), meaning that it represents the proportion of feature vectors that are classified correctly. The other cells of the matrix indicate false discovery rates (i.e., r − = |F + |/(|T + | + |F + |)) representing the rate of feature vectors that are misclassified as another gait event.

Gel Electrodes
Textile Electrodes  Subject  HS  TO  RG  HS  TO  RG  Same-leg event prediction  S3  15  15  59  15  15  59  S4  15  15  60  15  15  60  S5  15  15  38  15  15  42  Contralateral-leg event prediction  S3  15  15  55  15  15  55  S4  15  15  57  15  15  54  S5  15  15  44  15 15 40 India) makes them less reliable than the textile alternative. The latter, for example, can exhibit improved performance in face of perspiration caused by hot weather, as this tends to increase their conductivity and thereby allow for more sensitivity to muscle activation signal. In future work, it is planned to perform further experiments outside of the laboratory to assess such factors, and evaluate robustness of the proposed textile electrodes during daily use in a developing world context.