Participants

Forty-six subjects participated in the study: 30 women and 16 men. This was a representative sample of healthy volunteers from a group of students and academic workers from various universities in Prague. The average age was 28.2 ± 7.9 years. The mean height of the subjects was 173.8 ± 9.9 cm, and the most common weight range was 65–70 kg. Half of the subjects identified their right lower limb as dominant, 11 subjects identified their left lower extremity, and 13 subjects had no lateral preference for lower extremity as dominant. Exclusion criteria were any musculoskeletal injury in the past 6 months, acute illness, or pain while performing the test. All participants signed a written informed consent which is institutionally archived. The study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the ethics committee of the Faculty of Physical Education and Sport, Charles University under reference number 167/2020.

Study design

The study design focused on evaluating participants’ performance in two functional tests, the Single Leg Squat Test (SLST) and the Step-Down Test (SDT), following a standardized procedure. Participants first signed an informed consent form and were provided with information about the purpose and nature of the experiment. They then received video instructions detailing the proper execution of the functional tests to ensure uniform guidance. During the tests, three physiotherapists simultaneously assessed participants’ performance, while a precisely positioned RGB camera recorded the sessions for further analysis. The physiotherapists rated 15 possible execution errors for each test and performed a binary classification for each error (present or absent), providing an objective assessment of the participants’ test performance.

Functional test parameters and description

Our study examines the performance of two functional tests, the Single Leg Squat Test and Step Down Test. The performance of the used tests followed the protocol developed by Thonnard [2] and McGovern [3]. Following an explanation and demonstration of the test, the subjects were instructed to repeat the test three times for each lower limb. They always returned to an upright standing position between repetitions. Initially, the overall impression must be assessed (balance, gross arm deviation, ability to perform the test), and evaluated as satisfactory to proceed.

At least one of the three repetitions, evaluated as satisfactory, was sufficient to continue the test. The investigator then marked the following criteria as positive or negative:

1. Trunk flexion (forward lean, lateral rotation, lateral flexion, thoracic rotation)
2. Pelvic posture (tilt, rotation)
3. Hip position (adduction, internal rotation)
4. Knee joint position (valgus knee, tremor)
5. Depth of squat (compared to the other side, orientation with T)

Because the tests were not used as clinical screening for screening purposes, the overall results were not scored positively or negatively, but the parameters of each test were compared between evaluators and OP measurements. Prior to the test, the participants were given the opportunity to try both movement tasks with their right and left lower limbs once. Two images were then displayed on a screen in front of the participant. The first slide showed the initial position of the movement, while the second showed the final position of the movement.

Motion capture setup

For video recording, we purposefully chose a regular RGB camera: a Logitech C920 webcam with a resolution of 1280 x 720 and a frame rate of 30 frames per second, and h264 compression. In order to prove that we could make such measurements with an ordinary camera. We set this webcam on a tripod at a height of 130 cm to the top edge, and at a distance of 4m from the subject.

The camera was pointed perpendicular to the subject so that it would capture motion in the frontal plane. The green screen was placed 0.5 m behind the subject, leaving the distance between the camera and the screen at 4.5 m. Thereby we had the maximum chance of picking up body segments. For our study, we use a green screen in the background only for future use of the video dataset with systems operating on a different principle. The models used by OpenPose are trained in real-life environments. Therefore, the use of the system can also be in an environment with any background provided that there are no other people in the image that the system can also detect.

We instructed the subject to perform the aforementioned functional tests as described in the previous section. The Step Down Test was performed on a 20 cm step. Before the test, the subject had the opportunity to try both movements. They were shown two images showing the start and finish positions of both tests for ease of comprehension. Following this, the actual tests were conducted in the order of SLST, followed by SDT. The recorded videos had a length of approximately 15 seconds. A frame of the record can be seen on the right side of the Fig 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. In the left part you can see the 25 keypoints (landmarks) model of the OpenPose.

In the middle part is an image with the skeleton rendered while performing the functional test without any error. In the right part, the loss of balance error is clearly visible.

https://doi.org/10.1371/journal.pone.0288279.g002

Expert visual analysis

Our study utilizes the proficient judgment of three physical therapists (PTs), each holding a Master’s degree in physiotherapy and a practical experience of five years at the time of the study. These experts carried out an Expert Visual Analysis from a vantage point approximately 4.5 meters behind the recording camera, mirroring the camera’s angle to scrutinize the participant’s execution of the Forward-Step-Down-Test (SDT).

The selected configuration mirrors an approach supported by an analogous study from the Israeli Physical Therapy Society [29]. This study established that a rater’s familiarity with the SDT significantly bolsters the agreement rate, thereby emphasizing the critical role that familiarity plays in ensuring the reliability of the test results. Moreover, the study found that the level of work experience did not influence the agreement rate. These key findings reinforce the reliability and validity of our use of the SDT as an evaluation tool, enhancing the credibility of our methodology and the anticipated outcomes of our study.

Each PT independently evaluated the presence of deviations from correct movement execution using an electronic questionnaire and a binary rating scale. The PTs’ scores were evaluated for their inter-rater reliability using Fleiss’s Kappa [30].

As a result, we obtained an expert evaluation, which served as a reference. We used this information for categorization in our classification algorithms. Although we may not obtain precise biomechanical motion data, we can still develop an expert semi-objective reference standard based on subjective data acquired in this manner.

However, it is essential to note the potential limitations of this approach. The binary rating scale, while offering simplicity, may not capture the nuanced differences in movement deviations. Moreover, the validity of this scale is predicated on the expertise and subjective judgment of the PTs. Future studies may benefit from using a more detailed rating scale or incorporating additional methods to enhance the validity and reliability of the visual analysis.

Signal processing and signal extraction

Each of one subject’s measurements resulted in a single video file containing both tests. In order to minimize distractions, the video image was further cropped to show only the subject and the green screen behind him. With OP, we were able to extract spatiotemporal information from the video files.

OP extracted a human segment model from each frame. The collection of human skeletons is used to illustrate human movement in the video. We grouped the skeletons into a table with a row for each skeleton. The skeleton is anatomically defined by a vector, in which each vector component represents the position of a vertex in the skeleton. Fig 2 illustrates the two-dimensional Cartesian coordinates of a skeleton. The 25 key points of the model are automatically determined by OP. In our previous work, we show which anatomical points correspond to the points in the 25 landmarks model [31].

Among the detected landmarks, we calculate a signal of the range of motion in three variations. Angles and distances are calculated from the created segmental body model, which is based on identified landmarks. Custom-made software programmed in Python was used to calculate angles and distances.

Feature definition.

Based on the definition of functional tests [2], we selected specific elements to describe the correctness of execution, please see Table 1. These key points correspond to the anatomical points, please see study [31].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. List of selected ranges of movements(ROM), from which features were then counted.

https://doi.org/10.1371/journal.pone.0288279.t001

In this section, we describe signal processing. We used OP for video processing to extract key body points for each time and create a time series. From this time series, we calculated all presented angles and relative distances as signals. We then used these signals to extract feature parameters such as minima, maxima, medians, and means. The resulting data are available in S1 Data in a readable format. This approach allowed us to analyze the subject’s movements and identify specific features that are suitable for further processing by machine learning algorithms.

We are operating under the assumption that if an expert observes a movement, the relevant information is necessarily captured in the video. Although the expert may not see precise angles or measurements, their experience allows them to determine whether or not errors are present in the movement. Therefore, our focus is not on obtaining a precise biomechanical description, but rather on acquiring data that can distinguish between the presence or absence of evaluated parameters, and investigating whether these derived parameters contain information regarding the correctness of execution. To accomplish this, we utilize machine learning algorithms.

Classification

We have developed a learning system whose function is to answer six research questions by determining whether a specific phenomenon occurred during the execution of the exercise—six binary answers:

1. Loss of balance
2. Gross arm deviation
3. Trunk movement: Forward lean
4. Depth of squat
5. Posture of the hip joint: Drop
6. Overall Impression: Tremor

To train the machine, we utilized a Python library called PyCaret. PyCaret is a high-performance Python library with low code that facilitates the comparison, training, evaluation, tuning, and deployment of machine-learning models. In this library, we are able to evaluate, compare and adjust standard algorithms of different machine learning algorithms in parallel on the basis of a given data set in an efficient and comparative manner.

We choose to utilize AdaBoost, a method of ensemble learning (also called “meta-learning”), to improve our binary classification performance. The AdaBoost algorithm takes an iterative approach to learn from the mistakes of weak classifiers and converted them into stronger ones. AdaBoost is a sequential learning algorithm. Successive models are generated sequentially and their errors are learned by their successors. By giving the mislabeled examples higher weights, this technique exploits the dependency between models. Just as humans learn from their mistakes and do their best to avoid making the same mistakes in the future, the Boosting algorithm attempts to create a stronger learner (predictive model) from the mistakes of several weaker ones. The purpose of boosting is to reduce the bias error that occurs when models are not able to identify relevant trends in the data. AdaBoost (Adaptive Boosting) is a popular boosting technique that aims to combine multiple weak classifiers into a single strong classifier. The definition of a weak classifier is that it performs better than random guessing but is still ineffective at classifying objects. We have implemented the AdaBoost algorithm by using ten poor decision trees, processed sequentially.

For the AdaBoost classifier design in this study, we partitioned the available data into training, validation, and testing sets, allocating 80%, 15%, and 5% respectively. We utilized a Decision Tree Classifier as the base classifier with 50 weak learners employed in the AdaBoost classifier. The learning rate was set to 0.1 and the maximum depth of the decision trees was limited to 3. The feature subset size was determined as the square root of the total number of features and we set the random state to 42 to ensure reproducibility. We initialized the weights of all data points in the training set to be equal, iteratively training a weak learner on the training data, calculating the weighted error of the weak learner on the training set, calculating the weight of the weak learner based on their performance, and updating the weights of the misclassified samples. After the specified number of iterations, the predictions of all weak learners were combined using their weights. Finally, we evaluated the final AdaBoost classifier on the validation and testing data to assess its performance. The data partitioning and parameter details allow other researchers to replicate the experimental setup and enhance the reproducibility of the results.

Agreement of expert visual analysis

In Section, we described the 15 categories that the three independent PTs evaluated. Table 2 shows Fleis’s kappa and the size of each group. Some errors were not observed among the selected participants at all. Such groups could not then be subject to classification. Groups that were suitable for classification are highlighted in bold in the Table 2.

By using the Kappa coefficient to assess agreement between raters, we established a reference standard for classification that divides the data into those that can be used—those with high agreement—and those that cannot be used due to insufficient agreement among the experts. We followed standard methods to ensure the validity and reliability of our approach. Specifically, we used well-established procedures for data collection and analysis, and we implemented rigorous quality control measures to ensure that the results were accurate and reproducible. Overall, our approach provides a robust framework for classifying data and enables us to identify reliable and valid results for further analysis.

Results of classification

Our results show that if two or more raters agree on the presence of error, and there is a sufficient amount of measured data with and without error, our method based on a single 2D camera produces results comparable to the raters themselves. In the event that even three physiotherapists do not agree on the correct result, then it is likely that the gold standard for machine learning methods cannot be established. Due to the fact that our experts were not in agreement on all measurement categories, we were forced to exclude some of the categories from the evaluation. As a consequence, the experts’ assessment is still a subjective view and can differ depending on their life experience and the weight they give to individual errors.

For classification, we selected only the datasets fulfilling the conditions of kappa coefficient and minimum sample size, see Table 2. These were two separate datasets. Five responses (functional test parameters) to the first dataset (SDT), which contained 15 independent variables, met the conditions. The second data set (SLST) contained 15 independent variables and 2 responses from our panel of expert examiners. We calculated a confusion matrix for each dependent variable.

Our focus was on sensitivity and specificity. It was quite difficult to find subjects who did the movement incorrectly in our sample since only healthy people were included in our sample, according to our experts. In the end, our database is imbalanced because we have included only healthy subjects, the majority of the results showed that the patients were able to perform the exercises correctly. Therefore, we have implemented conventional methods in order to deal with the data imbalance. In this article, we wish to emphasize that since the purpose of the article is to prove the applicability of the method, we have chosen to present a number of cases in which a balance can be illustrated in the analysis. We are therefore very concerned with the sensitivity and specificity of our results in the results section. According to these two metrics, it can be seen that the data we finally used is balanced on the one hand, and on the other hand, they show that our method is able to detect any error by the subject with an average probability of 0.68 (Sensitivity) and that its prediction reliability (i.e. error reporting) is 0.8 (Specificity). The results are displayed in the Table 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Summary of the AdaBoost classification for the two exercises SDT and SLST.

In the case of the SDT, it presents five dependent outputs, while for the SLST, two variables are dependent. Dependent variables consist of binary classifications. The quality of classification is determined by the accuracy, specificity, and sensitivity of the classification. The sample is 46, for details, see Table 2.

https://doi.org/10.1371/journal.pone.0288279.t003

We used the AdaBoost Classification algorithm for our data analysis, which is a powerful ensemble learning tool for binary or multiclass classification problems. It integrates many weak classifiers to construct a robust classifier that provides accurate and exhaustive data analysis insights. Our approach trained weak classifiers, such as decision trees, on the weighted training set at each iteration to reduce classification error. We adjusted the instance weights by assigning incorrectly classified cases with greater weights and correctly classified instances with smaller weights, ensuring that subsequent weak classifiers focused on increasingly challenging events. The robust classifier produced predictions by weighing the votes of all weak classifiers, with the final forecast based on the class with the highest weighted votes. The AdaBoost algorithm is effective at delivering precise data analysis insights by leveraging the skills of several weak classifiers and focusing on difficult situations during training, resulting in a generalizable model for new, unobserved data.