Flexible and Wearable GRF and EMG Sensors Enabled Locomotion Mode Recognition for IoHT Based In-home Rehabilitation

— Benefiting from the development of the Internet of Healthcare Things (IoHT) in recent years, locomotion mode recognition using wearable sensors plays an important role in the field of in-home rehabilitation. In this paper, a smart sensing system utilizing flexible electromyography (EMG) sensors and ground reaction force (GRF) sensors for locomotion mode recognition is presented, together with its use under the IoHT architecture. EMG and GRF information from ten healthy subjects in five common locomotion modes in daily life were collected, analyzed, and then transmitted to remote end terminals (e.g., personal computers). The data analysis process was implemented with machine learning techniques (Support Vector Machine), through which the locomotion modes were determined with a high accuracy of 96.38%. This article demonstrates a feasible means for accurate locomotion mode recognition by combining wearable sensing techniques and the machine learning algorithm, potentially advancing the development for IoHT based in-home rehabilitation.


INTRODUCTION
It is of great significance to monitor the physical information, such as gait and muscle condition, of patients recovering from injuries or diseases under different locomotion modes to make a proper rehabilitation plan and evaluate the rehabilitation effect [1][2][3][4][5]. Tradition scenarios to obtain these data are mainly located in hospitals or therapy institutions [6,7]. However, patients may suffer pressure when being monitored, resulting in low quality data, conversely influencing the evaluation of the rehabilitation effect [5].
In recent years, with the fast development of wearable electronics and the Internet of Healthcare Things (IoHT) techniques, in-home rehabilitation has emerged quickly and gained worldwide attention, as in-home rehabilitation not only alleviates patients' undesired stress but also is capable of collecting high volumes of useful data by conducting long-term monitoring [8][9][10]. For in-home rehabilitation, locomotion mode recognition is important, which is closely related to the interpretation of body signals picked up by wearable systems [5,[11][12][13]. Hence, locomotion modes are must-have information delivered to hospitals under an IoHT architecture.
A locomotion mode recognition system generally consists of two parts: a data acquisition front-end and a recognition algorithm back-end. For the former, wearable sensors, including kinematic sensors (IMUs) [14][15][16], kinetic sensors (interaction force or ground reaction force sensors) [17], and electromyography (EMG) sensors [16][17][18][19] are widely used. For the latter, machine learning is frequently used for analysis of the physical information received from the front-end. For example, in [19], EMGs from nine muscles and six GRF measurements from load cells were collected, and Support Vector Machine (SVM) and Linear Discriminant Analysis (LDA) were used to identify five locomotion modes continuously with an accuracy no lower than 93.44%. In [16], seven EMG and two accelerometer sensors were used, and five locomotion modes were classified by SVM and LDA, with an accuracy of 95.2%. In [17], EMG, GRF, interaction force and position sensors were integrated to classify five locomotion modes with Bayesian LDA and reached an accuracy of 96.1%.
However, the successful broad use of locomotion mode recognition for in-home rehabilitation has not been reported yet. A possible explanation is that the utilization of multiple sensors not only increases the overall component cost, circuitry complexity, and energy consumption, which are undesirable factors for electronic products, but also brings users inconvenience when used. To address this, in this article, we present a smart locomotion mode recognition system with a high experimentally demonstrated classification accuracy of 96.38% by utilizing only two EMG and two GRF flexible sensors and a machine learning algorithm. Compared to the previous studies [14][15][16][17][18][19], the developed technique benefits from simplicity by smartly designing measurement locations and selecting highly correlated features for the machine learning algorithm while maintaining a high detection accuracy.

B. Flexible Sensors and Readout Architecture
The setup of front-end system is conceptually depicted in Fig.1. The flexible EMG sensor is a 10 mm-circular gelled electrode utilizing electrolytic gel as the interface between the skin and the metallic part of the sensor, whose composition is silver-silver-chloride (Ag-AgCl) [20]. Three electrodes were used to detect the EMG signal for a muscle with bipolar electrodes attached to the surface of the muscle and a reference one attached to the ankle [21]. Tibialis Anterior (TA) and Soleus (SL) were selected as the detected muscles and therefore six electrodes were used for EMG detection. Two flexible Force Sensitive Resistors (FSR) were attached to the flat area of the heel and hallux to collect the GRF signal.
The EMG and GRF signals were read out, preprocessed, and transmitted to a PC, as shown in Fig. 2. The raw EMG generated by the muscle was amplified by differential amplifiers to suppress the common mode input. An analog-to-digital convertor (ADC) and bandpass digital filtering (10 Hz-500 Hz [20][21][22]) was followed by the embedded MCU. For GRF signals, the resistance change of the FSR was linearly converted to a voltage change by a proportional operational amplifier and then converted to a digital signal by an ADC. Both the digital EMG and GRF signals were then conveyed to two Bluetooth devices (Huicheng, China) by the Serial Protocol and transmitted wirelessly via Bluetooth to PC. The major parameters of the front-end data detection system are summarized in Table I.

C. Locomotion Mode Recognition Algorithms
To recognize the locomotion modes of the user, a recognition algorithm using EMG and GRF signals from the front-end has been developed through three steps, as shown in Fig.3. Firstly, the EMG and GRF signals are preprocessed. All the datapoints are synchronized by timestamp. Normalization is performed to reduce the variability of each stride [23,24], and the signal sequence is segmented into each stride by a binarized GRF signal. Secondly, twenty-one statistical features, like the time domain features of EMG [25] and the force/time information of GRF, are extracted from the segmented EMG and GRF sequence in each stride. Finally, the extracted features were used to train the machine learning model, i.e., the SVM model, and then the model was used to predict the locomotion mode. A 10-fold cross validation was used to evaluate the accuracy of the trained model [26]. The above procedure was carried out in the MATLAB software and the classified locomotion mode together with the EMG and GRF profiles can be displayed in this software.

III. RESULTS AND DISCUSSION
The constructed wearable system is employed for recognizing users' different locomotion modes. A typical profile of the EMG amplitude and normalized GRF stress values under different locomotion modes is demonstrated in Fig.4. This figure demonstrates that the EMG and GRF profiles vary from each other under different locomotion modes. For example, when walking on level ground, the main activities of SL occur at about 50%-60% of the gait, and the heel will contact the ground prior to the hallux, as shown in Fig.4 (a), (f). In contrast, Fig.4 (e), (j) shows that SL is most active at about 10% of the stride, and  conversely, the hallux contacts the ground first when walking downstairs. This is because the front foot must reach out first, and the SL should contract in order to maintain balance in such cases. The profiles in Fig.4 indicate that the muscle activities and gait information were different in each mode in order to adapt to the relevant terrain. However, the profiles are sensitive to many uncontrollable factors, such as changes in the bodily conditions of the user. It would be difficult to recognize different locomotion modes by simply evaluating a few metrics with some fixed thresholds. Therefore, machine learning method is applied to comprehensively evaluate the non-linear changes of the EMG and GRF inputs to obtain robust recognition Here, an SVM model was trained and tested with extracted features. Fig.5 shows the confusion matrix of the model. An average accuracy and variance of 96.38% and 1.08% was obtained by the trained model. This model achieved the best recognition performance in recognizing SD, with the highest accuracy of 99.64% among all modes. This is because the EMG and GRF profiles of SD are rather unique and thus are easier to classify. However, the performance when recognizing some modes of the model, like distinguishing LW from RD, is relatively lower and less robust, with an error rate at 5.04% and a variance of 3.49%. A possible explanation for this is that, in some particular cases, like accidentally stepping out with an unbalanced stride, the human body will adjust naturally to avoid falling down, resulting in unusual changes in the EMG and GRF patterns and possible misjudgment of the classifier. This phenomenon (i.e., the error rate of recognizing RD into LW) is higher than the overall error rate and is also found in relevant works [27][28][29] (a comparison is given in Table II). A potential solution is to enlarge the volume of the training dataset. Fig. 6 depicts the application of the system developed in this paper under an IoHT architecture. The EMG and GRF information for people who need rehabilitation are detected and processed by front-end detectors; then, the signals are analyzed, and locomotion modes are recognized in end terminals, such as personal computers or smartphones. Finally, the integrated information is transmitted to remote end terminals. By monitoring changes in the locomotion modes and EMG and GRF data of the user, further analyses can be applied for various purposes, such as disease diagnosis or rehabilitation guidance. The developed IoHT system enables related individuals or institutions to monitor medical data and evaluate therapy performance with the patients at home or in other daily scenarios, which potentially reduces the burdens of both the therapists and the patients and broadens the interaction methods in the rehabilitation field.

IV. CONCLUSIONS
Locomotion modes recognition is of significance for inhome rehabilitation. The work presented here demonstrates the design and evaluation of a locomotion mode recognition system utilizing flexible sensors and machine learning technique. A recognition accuracy of 96.38% is obtained by merely utilizing two EMG and two GRF sensors, which offers a convenient and effective means to reach robust identification accuracy with lower complexity and energy consumption for IoHT based inhome rehabilitation.