1. Introduction
According to the WHO Global Health Estimates, cerebrovascular accidents such as stroke are the second leading cause of death and the third leading cause of disability worldwide [
1]. Amongst other pathologies, walking dysfunction is a major problem for many subjects who have suffered a stroke [
2,
3]. Furthermore, walking dysfunctions can lead to falls and restrict the patient in performing the activities of daily living. Thus, the improvement of independent walking is a primary goal of stroke rehabilitation [
4]. Besides conventional physiotherapy, robot-assisted gait training is used to counteract gait disorders. In a clinical routine, robots such as the Lokomat (Hocoma, Volketswil, Switzerland), Gait Trainer GT II (Reha Stim, Berlin, Germany), Lyra (Thera Trainer, Hochdorf, Germany) and G-EO (Reha Technology AG, Olten, Switzerland) are used. The benefits of these robots are twofold; the bodyweight support system of the robot carries the weight of the patient, and the gait pattern is induced by a robot specific strategy which leads to physical relief for the therapists. Additionally, patients who receive electromechanical-assisted gait training in combination with physiotherapy after stroke are more likely to achieve independent walking than patients receiving gait training without these devices [
5]. Besides the positive effects of robot-assisted gait training, there are also drawbacks of this technology. The bodyweight support system partially inhibits muscle activity [
6], and limited degrees of freedom of leg and pelvis movement can lead to changes in the naturally occurring muscle activation patterns [
7]. These drawbacks are tackled by developments of the aforementioned devices. Robot specific strategies, such as the biofeedback mode of the Lyra, the active-assistive/active mode of the G-EO or the alteration of the guidance force and FreeD mode of Lokomat, aim to improve the rehabilitation by allowing the subject to actively influence the robot-assisted gait training using their remaining muscle activity. Another approach to enhance rehabilitation is the combination of robot-assisted gait training and functional electrical stimulation (FES). The combination of these techniques is termed a hybrid robotic rehabilitation system, which has been shown to be more effective than robot-assisted gait training alone [
8]. Approaches to combine systems like Lokomat or the GaitTrainer GT II with FES were promising [
9,
10]. The G-EO, and the RT600 (Restorative Therapies, Baltimore, USA) system are commercially available hybrid robotic rehabilitation systems which can provide additional FES treatment during gait training. However, most of these systems depend on information (gait events) provided by the robot [
10,
11], focus on stimulating with intramuscular electrodes [
12], or focus on foot drop with a limited number of stimulated muscles [
9]. In order to extend the application and the use of robot-assisted rehabilitation, especially with FES, a novel algorithm was developed. In perspective, the aim is to combine inertial measurement units (IMUs) and robot-assisted gait training to trigger FES [
13]. This approach was feasible in a clinical setting, and was tested in a first step with a healthy subject. Thus, the objective of this study was to examine the robustness of the algorithm when used in a clinical setting with stroke patients, and to implement potential improvements.
3. Results
The following section visualizes the results of the gait event detection algorithm. In
Figure 9, the detection rates of the Lyra measurements can be seen, while
Figure 10 represents the results of the corresponding type 1 errors. The detection rates represent the percentage of correctly detected steps which could be used for a future trigger of functional electrical stimulation. In contrast to that, type 1 errors represent incorrect detected steps which could be a potentially trigger for hazardous electrical stimulation. Left and right inertial measurement units were analyzed independently from each other.
With the exception of two analyzed selection windows at the right sensor of the 1st subject, with detection rates of 70.7% and 72%, the detection rates for the left and right sensor were above 80%. Type 1 errors from the left sensors were 0% for fifteen out of twenty of the analyzed windows; the remaining five windows suffered type 1 errors below 7.5%. The right sensors suffered type 1 errors below 5%. The mean detection rate of all measurements of the stroke patients was 95.8% ± 7.5%. For type 1 errors, the mean value of all patient data was 1.0% ± 2%. One of the measurements (see
Figure 9 and
Figure 10) could not be analyzed (N.A.), as the sensor had to be removed during the measurement due to discomfort reported by the patient.
The detection rates of the Lokomat measurement are displayed in
Figure 11, while
Figure 12 represents the corresponding type 1 errors.
For the Lokomat measurements, all detection rates were above 90%, whereas twenty-three out of thirty-two analyzed windows reached a detection value of 100%, which means that every recorded step within those windows could be detected by the algorithm. For twenty-five analyzed windows, type 1 errors of 0% were achieved, five windows stayed below 3% and the remaining two windows had type 1 errors of below 11%. For the Lokomat, the mean detection rate of all patients was 98.7% ± 2.6%, and the according mean type 1 error was 0.9% ± 2.3%.
4. Discussion
Within this study, the robustness and feasibility of a gait event detection algorithm for robot-assisted gait training was demonstrated in a clinical setting with patients who suffered a stroke. The patients executed their therapy normally and did not face any disadvantages in the rehabilitation process due to participating in the study. Based on the presented results, the gait event detection algorithm is robust against patient specific movements and the lack of voluntary movements which might influence the behavior of the data during robot-assisted gait training. The newly introduced condition (
Section 2.5) is sensitive towards the detection of the Initial Contact, and enables the algorithm to neglect the unwanted detection of the Initial Contact due to the jittering of the jerk (
Figure 7). Despite introducing an additional technology (IMUs) to the therapists and the patient, no problems in the application of these devices occurred. The therapist and the patients showed great acceptance and interest in the experimental setup and its potential use within the therapy.
Some subjects were not familiar with robot-assisted gait training, which led to a shorter period of training time. One reason was that the setup time of the gait trainer took longer when the patient was not familiar with the procedure. In addition, longer explanatory conversations were necessary to explain the rehabilitation technique to the patients. Furthermore, exhaustion and the daily constitution of the patients affected the duration of the training. As a result, some analyses were limited to fewer analyzed windows compared to others.
Comparing the results of the detection rates of the Lyra (
Figure 9) to the previously conducted study with a healthy adult [
13], the detection rates are similar. The same applies to the results achieved with the Lokomat (
Figure 11). The type 1 errors for the stroke patients in both gait trainers (
Figure 10 and
Figure 12) are in a similar range compared to the type 1 errors of a healthy adult as well. Higher type 1 errors of around 11%, as reported in subject one in
Figure 12, may have various causes. The sensor might have loosened during the training, causing unwanted sensor behavior. Another reason could be that the settings of the exoskeleton caused an unrecognizable gait cycle, which lead to the improper detection of the gait events and resulting type 1 errors. As high type 1 errors may cause hazardous stimulations, these potential flaws of the experimental setup must be further investigated. Nevertheless, the type 1 errors were low for the majority of the analyzed measurements, providing good results for application in a clinical setting.
As mentioned in
Section 2.6, type 2 errors were not reported, because steps that are not detected by the algorithm cannot trigger a stimulation and are therefore not considered hazardous. However, high type 2 errors would influence the percentage of the detection rate and might lead to an inadequate stimulation rate. Thus, type 2 errors should be taken into consideration for the development of the prototype.
The subjects who participated in the study were in their early stage of rehabilitation and new to robot-assisted gait training (
Section 2.1). The robots used guide and limit the gait according to their robot specific strategy. Despite the guided movement, the level of gait abnormality might influence the gait pattern, and a hemiplegic leg might be a bigger disruptive factor than a non-hemiplegic leg. In order to evaluate this, a bigger patient population with stricter inclusion criteria should be recruited and analysed for future studies.
The proposed algorithm provides gait-event detection for level-ground walking during robot-assisted gait training. Other gait patterns, such as backwards walking, stairs and slope climbing, must be taken into consideration if the intended use of the algorithm should cover these aspects. So far, gait patterns other than level-ground walking have not been the subject of our research, but might be considered in the future development.
Approaches using techniques like LSTM-DNN (long short-term memory-deep neural network) have reached detection rates of up to 95.1% during level-ground walking on a treadmill [
20]. This approach uses the acceleration data from three sensors per leg, located on designated positions (instep, calf and thigh), to detect two gait phase events. Other approaches using deep neural networks on optical motion capture data for the automatic real-time gait event detection reached detection rates of 99% (for foot-contact) and 95% (for foot off) [
21]. Approaches using one sensor on the lower back [
22], thus minimizing the setup time, do not seem applicable for our purpose, as the bodyweight support system of the gait trainer covers the lower back and limits the movement of the upper body of the subject during the therapy. Comparing the results of our study to the above-mentioned approaches, we achieved similar detection rates using a lightweight, easy approach for practical application during clinical routine. Nevertheless, further studies with a higher number of subjects need to be conducted to prove the results achieved within this study.
As mentioned in
Section 2.4, the current system provides a simulated, faultless real-time Bluetooth connection by processing each stored sample individually and sequentially. Real Bluetooth connections may have a delay in transmition. Thus, investigations towards the development of a real Bluetooth connection and their potential problems will be part of the future research. The consideration of the delay of the whole measurement chain will be necessary to evaluate the development of the aforementioned prototype.
The implemented error handling methods to prevent algorithm failure cover a variety of potential errors. Nevertheless, a variation of the experimental setup, such as very high training velocities, could create a combination of events which have not been investigated so far. Despite the adaptability of the algorithm, more data, especially with non-tested velocities, should be recorded in order to prove the effectiveness of the implemented methods and to develop more processing methods to prevent algorithm failure.
In order to be able to generate a repeatable and comparable analysing process, a minimum number of two windows were analysed for each subject. As the shortest usable recording time of a subject was below eight minutes (the first subject in
Figure 11 and
Figure 12), a time of 3.5 min was chosen for the length of the windows. Robot-assisted gait training provides a repetitive movement. Thus, a longer or shorter time window should not influence the overall detection rate. However, it could happen that a specific movement of a subject, which might influence the data, is within or outside a window and could thus result in changes of the detection rate. Thus, modification of the window length might influence the detection rates, and should be considered in future analysis of the overall detection rate.
Overall, the results are promising, and the application of the algorithm seems feasible and robust when using it in a in a clinical setting for stroke patients. The algorithm is lightweight, and enables an easy setup for usage during a clinical routine. As a next step, the proposed algorithm should be used as a trigger for functional electrical stimulation, with the aim to further enhance the rehabilitation process of stroke patients. Additionally, further processing methods for algorithm failure must be included in order to ensure safe usage with electrical stimulation. In order to establish a prototype of the concept to combine robot-assisted gait training and IMUs to trigger FES, as proposed in [
13], further investigations and measurements must be conducted in order to optimize the algorithm and to synchronize the gait events with the stimulation.