1. Introduction
A survey revealed that more than 80% of elderly people aged 75 years or older were encumbered by degenerative arthritis [
1]. Furthermore, another study reported that one of every 100 people had rheumatoid arthritis [
2]. According to a report by the Agency for Healthcare Research and Quality, more than 600,000 Americans undergo total knee replacement (TKR) every year, while more than 15,000 people undergo TKR in Taiwan each year [
3,
4]. Osteoarthritis is the most common reason for knee replacement operation in the USA [
3]. The success of the replacement highly depends on the presurgery assessment and continuing post-surgery rehabilitation. Physical therapy is an essential part of rehabilitation after TKR. Regular rehabilitation is crucial for patients to adapt to the replaced artificial joint and gradually regain their mobility [
5].
According to a hospital’s tracking report, some patients after TKR had knees that remained swollen or even deteriorated when the patients returned to the hospital for inpatient services [
4]. Whether the surgical procedure was not performed properly and whether the patients lacked proper rehabilitation after returning home for recuperation are important clinical issues that concern orthopedists [
6]. Therefore, in this paper, we propose an effective method that integrates a designed sensor device, an inertial measurement unit (IMU), and an Android smartphone for monitoring the progress of rehabilitation after TKR. Furthermore, a software app was designed for long-term monitoring of the effect of rehabilitation by orthopedists and patients.
With the advancement of electronic technologies and their integration with wearable devices, the underlying systems have been extensively applied in the field of medicine, especially for tracking rehabilitation actions. To assess the evidence of wearable devices supporting their efficacy in assisting rehabilitation after total hip replacement and TKR, a review by Bahadori et al. [
7] found that both accelerometer and gyroscope were used in five studies. To objectively assess the lengthy process of rehabilitation for cerebrovascular diseases (e.g., stroke), Friedman et al. [
8] used a wearable sensor device (magnetometer) to precisely monitor daily use of the wrists and fingers. Mariani et al. [
9] attached wireless inertial sensors on shoes to estimate the heel and toe clearance for gait analysis. They proposed independent 2-D and 3-D models for motion detection and three other models for estimating gait rehabilitation information from patients without them being confined to a specific location and experimental environment for data collection.
Inertial measurement units (IMUs) have been widely applied to detect human postures and gaits in recent years. As a result of their low cost, ease of use, and low weight, sensors can be easily integrated with wearable devices and the Internet of things for various applications. IMUs can be used in posture identification for applications in sports and healthcare.
Van Der Straaten et al. [
10] gave systematic review to investigate the application of inertial sensor systems and kinematics obtained from systems in their study and to provide assessment to people with knee osteoarthritis and TKR. Kontadakis et al. [
11] introduced a gamified rehabilitation platform consisting of a mobile game and an IMU placed on a lower limb in order to capture its orientation in space in real-time for patients undergone TKR. Jones et al. [
6] attached IMUs on the lower limbs to objectively distinguish four rehabilitation exercises that were prescribed to osteoarthritis patients following TKR. Seel et al. [
12] used an IMU for joint axis and position identification, and for flexion/extension joint angle measurement. Bakhshi et al. [
13] developed two IMUs that were mounted on the upper leg and the lower leg to measure the knee joint angle.
Evaluation of gait and body motion disorder is relevant to fall risk assessment after knee-joint surgery [
14]. Huang et al. [
14] used three accelerometers (attached to both wrists and the chest) and applied the signal magnitude vector (SMV) and a type-2 fuzzy system for fall detection. Hsu et al. [
15] presented an automatic gait analysis algorithm that can automatically obtain acceleration and angular velocity by using one accelerometer and two gyroscopes. Triaxial accelerometers were used to detect falls events [
16,
17]. They presented a peak-value detection algorithm that can effectively discriminate the start and end times of each gait sequence. Furthermore, they proposed an algorithm that can improve the detection accuracy and quantify walking behavior under irregular movement. Mei et al. [
18] used acceleration sensor signals to determine gait event. Teufl et al. [
19] reported step width measurement based on IMUs and achieved valid results for 3D gait analysis. Müller et al. [
20] used a Kinect sensor to evaluate body motion and assess gait. Lorenzi et al. [
21] presented a mobile healthcare device to monitor human motion disorders. Gholami et al. [
22] designed a Kinect system to assess gait parameters in multiple sclerosis patients. Kun et al. [
23] developed a sensor set and an algorithm to estimate knee-joint kinematics. However, the captured signals from the aforementioned experiments [
23] cannot be processed online.
Data analysis and motion classification based on posture detection can help researchers more clearly understand users’ motion behavior. Some well-known methods such as neural networks, fuzzy modeling, and data mining have been proposed for analyzing motion patterns. Milosevic et al. [
24] used self-organizing maps for visualizing trunk muscle synergies during sitting perturbations. Han et al. [
25] used a single IMU to detect normal and abnormal gait phases. Hanakova et al. [
26] evaluated complex movements of the arm during walking based on gyroscope data and an angle–angle diagram. They also compared the results with those of the range of motion (ROM) method.
From the literature survey, we discovered that there remained problems to be solved in previous studies on rehabilitation monitoring after TKR. Therefore, the developed sensor-based system is aimed to fulfill three objectives: (1) monitoring whether the TKR patients followed the orthopedist’s rehabilitation instructions at home, (2) recording the duration of each rehabilitation session, and (3) determining the extent to which a patient’s knee can flex in each rehabilitation course. To fulfill these objectives, the developed sensor devices had to be reliable, user friendly, and easy to use and had to enable TKR patients to achieve rehabilitation effects at home similar to those achieved using professional equipment in a hospital.
This paper is organized as follows. In
Section 2, the proposed method and hardware design are described, and the procedure for calculating the equivalent range of motion (ROM) is detailed. The way to monitor rehabilitation progress is illustrated in
Section 3. The experimental results and discussion are provided in
Section 4. The conclusions and future work are presented in the
Section 5.
3. Monitoring Rehabilitation Progress
During a rehabilitation course, the users’ knee motion angle varies over time. The distribution of the swing angles may range, for example, from 60° to 180°. Even if we were to record all motion angles, there is no simple method to quantitatively measure the effect in each rehabilitation course. To resolve this problem, we applied Fuzzy c-means (FCM) to identify the centroid of the acceleration signals so that an equivalent ROM can be calculated to represent the effect of a rehabilitation course.
3.1. Equivalent Angles of Knee Motion from FCM
FCM is one of the commonly used machine learning methods that can softly partition data into the predetermined number of clusters [
31]. A datum can be classified into any of the clusters with a membership degree between 0 and 1 under the constraint that the sum of membership degrees should be equal to 1. FCM was applied to calculate the equivalent ROM for the swing angles of rehabilitation. When a pair of the developed sensor devices were worn on the thigh and ankle, as shown in
Figure 9, the angle between the thigh and the shank was 180° –
θ, where
θ is the angle between the shank and the ground.
The triaxial accelerometer is highly sensitive, therefore any vibration or other disturbance during the rehabilitation course causes an abnormal reading from the device. Therefore, before determining the equivalent angles, gravity was removed and the Kalman filter was then applied to preprocess the received signals from the sensor devices. Using the Kalman filter to preprocess the raw acceleration signals can suppress disturbances and smooth them for further analysis. The equivalent ROM was calculated from the centroids G1 and G2, as shown in
Figure 10. The steps to calculate the equivalent swing angles are listed as follows:
Step 1: Record the signals from both sensors 1 and 2. The swing effects from the left and right directions were discarded, and only the acceleration signals from Xg and Yg were considered.
Step 2: Use FCM to cluster the signals from each sensor into three groups, represented by fcm1i and fcm2j for i, j = 1, 2, 3.
Step 3: Find the centroid from each sensor device and represent the pair as (G1, G2).
Step 4: Perform basic operations on inverse trigonometric functions to calculate the equivalent ROM from the centroid pair (G1, G2) as follows:
3.2. Monitoring the Effect of Rehabilitation
An Android smartphone was used to receive and record signals transmitted from the developed sensor devices. The users needed to input some basic information, such as name, age, gender, and the preferred animation type (boat, cow, or car) at the first time of use. The designs for the animation types were based on the fact that most elderly people living in the vicinity of the hospital were fishermen, farmers, and retirees. After this information was input, the smartphone was paired with the sensor devices via Bluetooth, as shown in the lower part of the start page. Once paired successfully, the smartphone was ready to receive and analyze the signals transmitted from the sensor devices. An orthopedist can simultaneously track and monitor a patient’s rehabilitation status. If a patient does not follow the instructions, the orthopedist can actively contact the patient to determine what the problems are.
The maximum swing angle is displayed sequentially in the lower part of the smartphone screen each time, as shown in
Figure 11. Properly recording the swing time is crucial for clustering the angles and evaluating whether regular rehabilitation is performed. The counter for each rehabilitation course is highly useful for monitoring recovery progress, as the orthopedist can refer to the counter to decide whether to adjust the rehabilitation course. The counter for each rehabilitation course is also displayed on the screen. We use a color bar to display the percentage of completion of the designated course. The selected animation type, for example a car, moves upon each swing. This approach encourages the patient to continue exercising; otherwise, the car stops moving.
To monitor the effect of rehabilitation, a line chart is used to display each swing angle on the screen, which represents a considerable improvement over the current rehabilitation system in which patients must wait before receiving an examination report. Furthermore, quantile plots are presented to display the swing angles in ascending order as well as to quantitatively compare whether there is noticeable progress after the designated course.
3.3. Quantile Plot for Swing Progress
To quantitatively monitor the long-term progress of rehabilitation, we used a quantile plot to display the swing angles on the app. After the completion of one exercise course, the swing angles were saved in the memory and displayed as a line plot in the app, as shown in
Figure 12a. The lower part of the screen in
Figure 12a showed that the subject performed 52 swings with an average angle of 82.11°. For validating the effect on rehabilitation, the swing angles were rearranged in ascending order. For illustration, two dotted line plots are compared in
Figure 12b, where the
X-axis represents the percentage of swing angles, and the Y-axis represents the corresponding angles. A line drawn perpendicular to 50% (f = 0.5) intersects with the bottom curve at an angle of 60° and with the upper curve at 75°. This implies that the first 50% of the swing angles were lower than 60° in the first rehabilitation course, whereas the leg flexion angle improved to 75° in the second rehabilitation course, indicating progress. By contrast, the swing angle remaining low, the line plots being flat, or the plots not having noticeable differences relative to the previous plots for a long time indicates either that the surgery was not completely successful or that the replaced knee joint gradually hardened. This may require immediate intervention by orthopedists to identify the problems.
5. Conclusions and Future Work
Monitoring whether TKR patients are rehabilitated after the surgery remains a major concern for orthopedists. Without continuing rehabilitation, full recovery is delayed and the weak knee joint may affect the mobility of the patients, resulting in an urgent need for new devices or methods to overcome these problems. In this paper, an effective method is proposed to resolve the problems using three approaches: (1) monitoring whether the TKR patients follow the rehabilitation instructions at home, (2) automatically recording the duration of the rehabilitation course, and (3) saving the flexion angles and monitoring the progress from each rehabilitation course.
The proposed sensor device has social benefits and advantages such as usability without spatiotemporal constraints, reduction of frequency returning to the hospital for inpatient services, saving medical expenses, and accuracy in monitoring the rehabilitation progress. The developed sensor devices can be easily worn on the thigh and ankle, and the proposed method can calculate the number of swings and the equivalent ROM from each rehabilitation course. This fulfills the second and third goals of this study. An app was designed to display the swing angles so that users can track the effect of rehabilitation. The orthopedist can also monitor the progress of rehabilitation, thereby fulfilling the first goal of this study. The experimental results show that the average absolute swing errors from the TKR subjects were between 1.65° and 3.27° and that the accuracies were between 98.09% and 96.16% at different angular speeds.
Although the developed sensor devices are small and lightweight, they must be placed into a shell, and Velcro and an elastic belt are required to wear them on the leg. The hardware is proposed to be modified into a chip in the future. Furthermore, the developed devices are under investigation for other medical applications, such as rehabilitation for frozen shoulder, measuring trembling in Parkinson’s disease, identifying gait and joint patterns during walking, evaluating medial/lateral load or possible excessive stress shielding growth, and sport applications such as pitching pattern identification and adjustment.