Inter-rater and intra-rater reliability of isotonic exercise monitoring device for measuring active knee extension

Subjects

Healthy volunteers were recruited to participate in the study. Eligible subjects were at least eighteen years of age, had full ROM in both knees and had no history of musculoskeletal or neurological problems affecting either leg. The research team advertised the details of the study at Prince of Songkla University, and all eligible subjects who contacted the research team were recruited to the study. All subjects cooperated well and provided written informed consent. The study protocol was approved by the institutional ethics committee of the Faculty of Medicine, Prince of Songkla University, and was in accordance with the Declaration of Helsinki (REC 60-284-11-1).

Study design

This was an experimental study. To determine the inter-rater reliability of the isotonic exercise monitoring device, sEMG and ROM measurements were obtained and compared with standard instruments: the Mobi6-6b wireless system and the computer vision method, respectively. For intra-rater reliability, the subjects were tested one day apart using the same protocol. In both experiments, the speed, time, and force of the exercises were controlled by the Con-Trex. All measurement devices were synchronized to the time alignment of the Con-Trex to accurately determine the period of knee extension and flexion. Experienced physical therapists monitored the experiment and were in charge of the Con-Trex to ensure both safety and reliability.

Standard instruments

We used three main instruments in this study to verify the results of the isotonic exercise monitoring device. The Con-Trex dynamometer, used in this experiment to control the exercises, is popular among exercise monitoring devices because it can monitor isokinetic, isometric, and isotonic activities. Previous studies have shown it has excellent reliability for knee exercises (Maffiuletti et al., 2007) with an ICC of 0.99 for knee extension and an ICC of 0.78–0.81 for knee flexion.

We used the Mobi6-6b wireless system, a wireless device with multiple channels and applications, to measure the sEMG in this study. A previous study (Sae Jong, de Herrera & Phukpattaranont, 2021) revealed that, when given a variety of sEMG measures for facial muscle signal detection, such as mean absolute value (MAV), root mean square (RMS), and frequency domain, the Mobi6-6b had 98% accuracy in syllable recognition.

We used computer vision as the standard for measuring knee range of motion. There are numerous applications employing computer vision for monitoring human activities such as sitting posture (Kulikajevas, Maskeliunas & Damaševičius, 2021), human gait (Yu et al., 2021), and foot recognition (Duangsoithong, Jaruenpunyasak & Garcia, 2019). One study showed that applying deep learning through a convolutional neural network (CNN) gave a high, 98% accuracy for recognizing the region of interest (ROI) of the foot in images, which could then be employed to monitor the foot during knee exercises.

Experimental procedures

The methods of this experiment consisted of two main steps: preparation and exercise, as shown in Fig. 2. To prepare for the isotonic knee exercises, each subject completed a series of warm-up exercises followed by 5 min of brisk walking (Kim et al., 2014) and then a 5-min rest period. After the rest period, the subject was seated on the Con-Trex station in a comfortable posture, and his/her back was strapped to the back rest for safety. The subject’s knee position was set to 90 degrees of flexion. Because of limited space and the distance between the camera and the subject, we tested only the subject’s right leg. Next, the lever arm of the Con-Trex was attached to the ankle. The skin beyond the VMO muscle was then cleaned with an alcohol pad to remove dirt and oil. This procedure additionally reduced electrode impedance mismatch (Toledo-Peral et al., 2018). The team of researchers (J.J. and K.S.) then attached the two sEMG electrodes to the skin just past the VMO muscle with the isotonic exercise monitoring device. Specifically, the two active electrodes were attached on the line between the anterior superior iliac spine and the joint space in front of the anterior border of the medial ligament, following SENIAM guidelines (De Luca, 1997; Hermens et al., 2000). The distance between the two electrodes was 20 mm. One reference electrode was placed on the C7 spinous process. Following electrode placement, the triaxial accelerometer of the isotonic exercise monitor was strapped around the right ankle below the edge of the Con-Trex’s lever arm to measure knee ROM. Finally, the researchers configured the parameters of the isotonic exercise monitoring device as shown in Table 1. The researchers also set up the video camera system 2 m away from the subject with a lateral view perpendicular to the subject.

For the exercise procedure, the one-repetition maximum (1-RM) of each subject was measured by having each subject perform just one isotonic knee extension exercise, followed by a 5-min rest period (Freitas de Salles et al., 2009; Lin et al., 2008). The subject then performed two isotonic knee exercise repetitions at 25% of his/her 1-RM to become familiar with the measurement and to check the sEMG signal and ROM measurements from the isotonic exercise monitoring device, followed by another 5-min rest period. The subject then performed six isotonic knee exercise repetitions using the Con-Trex device, with the isotonic exercise monitor recording the sEMG and ROM of the knee, and the computer vision program also measuring the ROM of the knee. Following this, the subject rested for another 5 min (Freitas de Salles et al., 2009; Lin et al., 2008), during which the researchers (J.J. and K.S.) replaced the sEMG cable of the Mobi6-6b device’s electrodes with those from the prior experiment, configuring the Mobi6-6b device parameters as illustrated in Table 1. The subject then performed six isotonic knee exercise repetitions using the Con-Trex device, with the Mobi6-6b device recording the sEMG signals. At the end of the experiment, the sEMG electrodes on the skin were removed, and the location of the electrode attachment was marked for the next experiment for each subject. All subjects returned 24 h later and performed the same experiment to evaluate the intra-rater reliability of the isotonic exercise monitoring device compared with standard protocols. All data were automatically collected by all instruments used in the experiment.

Data analysis

The motor unit action potential (MUAP) is the combination of all muscle fiber action potentials from a single motor unit (Raez, Hussain & Mohd-Yasin, 2006). In this study, the MUAP was detected by a skin surface electrode positioned on the VMO muscle. At the electrode, the raw sEMG was detected and amplified. Differential amplifiers are often utilized as first stage amplifiers. Before collection, low- or high-frequency noises and other signal noises were filtered from the signal as demonstrated in Fig. 3.

The raw sEMG signal from the isotonic exercise monitoring device was first amplified 1,500-fold and then filtered by the bandpass filter circuits at 10–500 Hz. The notch filter circuits subsequently reduced the 50 Hz as powerline noise (Akwei-Sekyere, 2015) in the sEMG. Next, this sEMG was sampled at a frequency of 1,000 Hz as part of the software used to gather data into the computer. The signal was also filtered with an infinite impulse response (IIR) bandpass filter at 10–500 Hz and then normalized to a standard value by dividing the maximal peak of sEMG during muscle contraction, according to the methods of prior studies (Biviá-Roig, Lisón & Sánchez-Zuriaga, 2019; Gagnat, Brændvik & Roeleveld, 2020). Time synchronization by Con-Trex was then used to partition this sEMG by knee extension periods. Finally, the MAV and RMS were extracted from the sEMG signal and calculated using the following two equations:

(1) $$MAV=\frac{1}{N}\sum _{i=1}^N\left|X_i\right|$$

(2) $$RMS=\sqrt{\frac{1}{N}\sum _{i=1}^N{X}_i^2}$$

Here, MAV is the mean absolute value of the sEMG signal, N is the number of sEMG samples, $X_i$ is the sEMG amplitude at the $i$ index, and RMS is the root mean square of the sEMG signal. The raw sEMG signal from the Mobi6-6b was amplified 1,000-fold and then transferred using Bluetooth technology. In the subsequent procedure, the computer received wireless sEMG data with a sample rate of 1,024 Hz, which was then filtered with an IIR bandpass filter at 10–500 Hz. Finally, the sEMG features were extracted using the same methods that were used with the isotonic exercise monitoring device.

To calculate ROM, as illustrated in Fig. 4, a triaxial accelerometer ( $\pm 2$ g range) of the isotonic exercise monitor was strapped around the right ankle of the subject and then sampled at 100 Hz. The roll ( ${\varphi }_a$) and pitch ( ${\theta }_a$) angles (Narkhede et al., 2021; Franco et al., 2021) were then calculated using the following equations:

(3) $${\varphi }_a=\arctan (a_y/a_z)$$

(4) $${\theta }_a=\arctan (\frac{-a_x}{a_y\sin \varphi +a_z\cos \varphi })$$

Here, $a_x$, $a_y$, and $a_z$ are represented linear accelerations determined along three orthogonal axes, respectively.

The ROM was measured from the starting angle of knee flexion to the final angle of knee extension based on the time-synchronization of Con-Trex recordings. During knee exercise, the subject was recorded using a digital camera (Canon EOS 550D) as part of the computer vision approach. The digital camera was also mounted on a tripod that was 1 m height and perpendicular to the subject. There are six primary ROM calculation processes based on convolution neural networks in object detection (Hakim & Fadhil, 2021). First, an image of the knee at its initial angle of flexion was cropped from the top left to the bottom right using sliding windows to find the located foot ROI in the image. According to the methods outlined in a previous study (Duangsoithong, Jaruenpunyasak & Garcia, 2019), a pre-trained convolutional neural networks (CNNs) model was then used to classify each cropped image to determine whether it contained a foot or not. If this CNNs model predicted the that image contained a foot, the ROI was stored in memory until the end of the last sliding cropped image. The algorithm then picked a single ROI from several overlapping entities, a process known as non-maxima suppression (Rothe, Guillaumin & Van Gool, 2015). Next, the centroid of the foot ROI and the fixed location of the knee joint was labelled by the team researcher (J.J.). In this experiment, the knee joint of the subject was fitted into the Con-Trex lever arm’s rotating point. The location of the camera and the Con-Trex both remained the same. The angle of the knee joint along the Y-axis was calculated using the ROI of the foot and the location of the knee joint, as shown in Eq. (5) and illustrated in Fig. 5. Then, the final knee extension degree was noted, as labeled by the Con-Trex. The ROM was then determined by subtracting the end degree of knee extension from the beginning angle of knee flexion. After calculating the range of motion, the video of the participant’s exercise was deleted immediately. Only the final ROM calculation of the participant was stored on the computer.

(5) $$\theta =\arctan (A/B)$$

Here, $\theta$ is the angle of the knee joint, A is the distance on the X-axis between the centroid of the foot ROI and the location of the knee joint, and B is the distance on the Y-axis between the centroid of the foot ROI and the location of the knee joint.

Statistical analysis

For the sample size calculation, we used the ICC based on two raters (k = 2) (Jerez-Mayorga et al., 2020), and the expected effect size to assume differences in EMG and ROM measurement in an isotonic exercise monitoring device compared with standard protocols was computed (Koo & Li, 2016). We estimated the power (1 − $\beta$) by defining the expected reliability (ICC, $\rho 1$) as 0.6, the significance level ( $\alpha$) as 0.05, and a power of 0.90 (Bujang & Baharum, 2017). The sample size estimate indicated that at least 20 subjects were required.

The baseline characteristics of participants were presented as means and standard deviations (SD). The mean value of both devices on each day was calculated to evaluate the inter-rater and intra-rater reliability using ICCs. Moreover, an absolute reliability was calculated with the standard error of measurement (SEM). The research team proposed a method for interpreting ICCs, which has been used in many published studies (Koo & Li, 2016). An ICC value higher than 0.90 indicates excellent reliability. ICC values between 0.75–0.90 and 0.50–0.75 indicate good and moderate reliability, respectively. Lastly, ICCs less than 0.5 indicate poor reliability. The Bland−Altman plot analysis was conducted to demonstrate the agreement between the isotonic exercise monitor and the standard measurement devices. Systematic bias was determined using the 95% CI of mean difference (Gerke, 2020). If zero is included in the 95% CI, no systematic bias may be inferred. A narrow 95% CI (LoA95) indicates more stability. We utilized all analyses using the R Program (Vienna, Austria) version 4.0.3.