1. Introduction
Hemophilia A and B are hereditary disorders characterized by a deficiency in clotting factors VIII and IX, respectively. This deficiency causes a chronic tendency to hemorrhage [
1]. The frequency of this disease is low, so hemophilia is known as a rare disease; for example, hemophilia A occurs in about one in every 5000–6000 live newborns and hemophilia B in one in 30,000 [
2]. However, despite the low incidence and prevalence, the cost associated with this disease is one of the most elevated, so it has a great impact on the health system [
3]. One of the most important and most increased morbidities in patients with hemophilia is bleeding occurrence in joints, known as hemarthrosis [
4]. This bleeding may occur due to a blow or spontaneously, due to friction during the natural movement of the joint. Without proper treatment, recurrent hemarthrosis causes hemophilic arthropathy [
5], which involves chronic pain and functional disability of the joint [
6]. In patients with hemophilia, bleeding occurs mostly in the knee (44%), elbow (25%), ankle (15%), shoulder (8%), hip (5%) and in other locations (3%).
Health benefits that involve physical activity for the general population are equally applicable to people with hemophilia. An adequate muscle tone avoids injuries and decreases the risk of bleeding in joints. Physical exercise is of great importance for the hemophilic population because it improves their quality of life [
7]. There are many studies that address the importance of physical training to improve health condition. They emphasize the World Health Organization recommendations [
8] for the general population and World Federation of Hemophilia recommendations for the hemophilic population [
7,
9,
10].
In particular, the use of video games for rehabilitation (exergaming) is having a positive impact on the patients’ attitudes towards training and has proven useful to enhance their strength, coordination and mobility [
11,
12]. Furthermore, motion capture (MoCap) sensors are being increasingly used for applications in medicine and in physical therapy, as these sensors are becoming readily available in the market and relatively inexpensive against other alternatives such as 3D optical MoCap systems [
13].
A popular choice of a MoCap sensor is Kinect, which has been widely (and successfully) used for rehabilitation in a wide range of medical applications [
14,
15,
16,
17,
18], such as post-stroke limb rehabilitation, elderly exercise monitoring and fall prevention [
19], range-of-motion (ROM) evaluation in patients with adhesive capsulitis [
20], balance and postural control assessment and training [
21,
22], virtual gyms for people with restricted mobility [
23], etc.
The validity of the first version of Kinect (V1) for clinical applications has been the object of study of several publications [
22,
24,
25]. Several studies [
26,
27] point out the limited accuracy of Kinect V1 against a clinical goniometer for joint angle measurement, although other recent studies have reported excellent agreement between their measurements [
20]. Kinect V2 has been reported as more accurate than V1 [
28,
29,
30] and has recently attracted attention to devise and evaluate a variety of exercises aimed at rehabilitation for patients with axial disorders [
31], ROM measurements for home rehabilitation [
32] and gait parameter measurement [
33].
Until now, only two studies have used MoCap techniques in patients with hemophilia: one for rehabilitation [
34] and the other for balance assessment [
35]. In both studies, a Nintendo Wii Balance Board was used. Therefore, the main objective of this work is to create a redistributable software tool, HemoKinect, that sets up a number of key exercises to be performed in front of Kinect V2. HemoKinect allows patients to perform them interactively either at health facilities or comfortably at their homes through a user-friendly graphical user interface (GUI). The specialist should be able to remotely receive and interpret feedback from the patients’ activity, via comprehensive periodical reports, to keep track of their progress and adjust their training.
3. Results
Example plots of each exercise for random patients are shown in
Figure 4. For the elbow exercises, the threshold variables
flexionThresholdAngle and
extensionThresholdAngle were set to 50
and 100
, respectively. For the knee flexion exercise, the variables
flexionThresholdAngle and
extensionThresholdAngle were set to 70
and 30
, respectively. These values were selected according to the doctors’ advice taking into account the mobility limitations of the patients due to arthropathies, while allowing the correct exercise count in controls and avoiding false positive detection (i.e., counting flexions by mistake when walking normally).
Five series of five repetitions per series were collected for each type of exercise and participant. We measured the achievement rate, i.e., the percentage of exercises correctly performed as judged by HemoKinect, for each type of exercise.
For elbow and knee exercises, a 100% of achievement rate was obtained for the control population. In the hemophilic group, the achievement rate for the elbow was also 100%.
Table 2 and
Table 3 show the rest of the achievement rates, for the control and the hemophilic population, respectively. A perfect performance was measured for the knee exercise for half of the patients, and the average achievement rates over all patients were high (85% for the left knee and 86% for the right knee), except for those who presented high levels of arthropathy in their knees, which did not allow them to reach the detection threshold in some repetitions. As mentioned in Algorithm 3, HemoKinect relies on the knee angle in combination with the hip center position elevation and descent as elements used to detect step climbs and descents. In this manner, step climbs were successfully counted independently on both feet in about 78% of the cases for the right foot step and 75% for the left one in the control population. The patients’ achievement rate falls approximately by 15% with respect to the controls’ rate.
The balance exercise was evaluated using only the hemophilic population. Statistics have been obtained by averaging the score per patient/difficulty level/cardinal direction. The results are shown in the box plots of
Figure 5.
Figure 5a shows that each participant obtained a different balance performance. Three patients presented an average performance below 50%, two patients a performance close to 65% and the remaining three patients close to 75%. As expected,
Figure 5b shows that the balance performance decreased as the difficulty level increased. Kruskal–Wallis analysis revealed that there were significant differences among the levels of difficulty (
p-value = 0.013). Multiple comparisons find differences (
p-value = 0.010) between the performance at Level 1 (median = 68.50 and interquartile range = [58.00–76.00]) and at Level 3 (60.00 [46.25–71.75]. However, no significant differences were found between Level 2 (65.50 [51.00–73.75]) and Level 1 or between Level 2 and Level 3.
The Kruskal-Wallis analysis also revealed that there existed significant differences in the balance performance as a function of the target direction (p-value < 0.001). Differences were found between the following directions: S (56.00 [43.75–67.00]) and W (73.50 [65.25–77.00]; p-value = 0.008), S and NW (70.00 [57.75–79.75]; p-value = 0.042), S and E (72.00 [57.25–85.50]; p-value = 0.004), N (56.00 [45.00–67.25]) and E (p-value = 0.005), as well as between N and W (p-value = 0.011).
4. Discussion
The experiments of elbow flexion/extension and knee flexion/extension with empirically tuned threshold angles have been able to successfully detect exercise repetitions for the control and hemophilic population. In the performed tests, an elbow flexion/extension success rate of 100% was achieved by both populations. The squats were successfully performed by all controls and most patients involved in the study. The accuracy of Kinect reveals great potential in the systematization of ROM measurements of multiple patients at the medical consultation. Additionally, by saving a record of the joint data, the specialist could detect potential postural injury risks and help prevent them.
There exists a wide range of commercial MoCap sensors that may be used for clinical applications as, for example: Orbbec Astra, RealSense R200, ZED Stereo Camera, RealSense F200, DUO mini lx, Leap Motion and Kinect V2. Kinect V2 presents a series of advantages over the alternative MoCap sensors: it can be used for full skeletal tracking and track multiple bodies simultaneously, which may allow parallel acquisition of data from several patients; it has a relatively low price; it supports a wide variety of software toolkits (Open-Frameworks, Processing, Unity3D, etc.) and languages (C#, C++, JavaScript, Java, etc.); and it has mature drivers and a well-documented, freely-accessible SDK.
Previous studies that used depth sensors for rehabilitation purposes have mainly used the previous version of Kinect (V1). There are mixed opinions on whether this system should be used for clinical evaluation or not. There are authors that report that angle accuracy is not enough for medical purposes [
26] and others that state that they can be acceptable for a rehabilitation tool [
14]. Kinect V2 possesses an improved depth sensor and higher resolution, together with the ability to track a larger number of bodies and joints per body. As a result, Kinect V2 becomes a valid alternative for clinical applications, as shown by some authors [
28]. In our experiments, Kinect V2 has been able to measure elbow and knee angles when the joints of interest are not occluded by any object. Occlusion has proven to be a limitation since the first Kinect V1 [
24] and continues to be a source of error in V2. This means that, in all tests, the patient must be placed facing the sensor and within its recommended range (0.5 m–4.5 m) and under uniform lighting.
The limitation when designing ankle exercises is the inaccuracy in the detection of the ankle position provided by Kinect V2. This was also reported by other authors [
28,
30]. Although a regular pattern is observed in successive ankle flexions and extensions, the positioning of joints that participate in the ankle movement (particularly the foot segment) is greatly distorted by noise. This fact makes all tested exercises based on the measured ankle angle for counting ankle flexions/extensions unreliable. This limitation was partially bypassed by designing an exercise that implies a step detection. The designed algorithm successfully counts climbs and descents, based on the hip center joint position and knee angle, implicitly forcing the adequate mobility of the ankle in each repetition. In some cases, the occlusion of the ankle joint by the step block may cause errors in the calculation of the knee angle, which may lead to misdetection of some repetitions. Assuming this limitation, the step algorithm can detect both right foot and left foot step climbs independently and has reached average performances between 64% and 78%, depending on the population.
Balance training in exergaming applications has been previously addressed in the literature. The first Kinect V1 has been reported to be effective in accurately characterizing changes in the COM and in flexion-extension movements of the lower limbs during balance training [
21]. For hemophilic patients’ rehabilitation, a Nintendo Wii Balance Board has been evaluated [
34]. However, the Wii relies on pressure distribution, while Kinect composes a 3D point cloud, and our algorithm can extract the actual COM position.
The balance experiment produces interesting results. As observed from
Figure 5a, the balance performance varies in each patient. These differences may be due to their physical condition and level of arthropathy. As expected, according to
Figure 5b, the general balance performance decreases with the increasing level of difficulty. However, the results for Levels 1 and 2, as well as for Levels 2 and 3 overlap significantly, and the differences are only obvious between Levels 1 and 3. This implies that the change between levels is not abrupt, but rather progressive. The rapid motor learning of the participants during the exercise execution allowed them to take on Level 3 even if some of them started struggling to reach some positions at Level 1.
For anatomical reasons, swaying to reach the player’s front target positions should be easier than swaying to reach the back ones, and this has been proven by multi-directional reach tests [
42,
43]. This generally agrees with our results, except for the N direction, as observed in
Figure 5c. The explanation for this inconsistent behavior is the fact that this was the first direction the patient was asked to reach. Therefore, their reaction time was higher than for the rest of the directions, and as a consequence, their performance decreased. Predictably, the S target direction obtains the lowest scores, as it is affected by the fear of falling factor [
43]. In terms of performance per cardinal direction, the Kruskal–Wallis analysis has proven that statistical differences are only found between directions that are not adjacent.
On the whole, the results of this pilot study are very promising. In the future, the amount of participants of the study will be substantially increased thanks to the distribution of the hardware and software, and statistical tests will be carried out for a long-term measurement campaign of HemoKinect against traditional rehabilitation techniques. HemoKinect is also able to measure the execution time of all exercises, and this could be exploited at advanced rehabilitation stages to encourage patients to speed up their exercises and improve their musculoskeletal health.
Limitations
In October 2017, it was announced that further development for Kinect V2 will be discontinued. Nevertheless, Microsoft will continue to provide support for the Kinect SDK and is working with Intel to provide an alternative [
44]. There are other sensors with body tracking SDKs on the market such as Intel’s RealSense [
45], VicoVR [
46] or Orbbec Astra [
47] that could be evaluated as alternatives to Kinect V2.
Kinect V2 applications already developed, such as HemoKinect, will continue to work, and clients will be able to use them without any issue, as long as they are in possession of Kinect V2 hardware and the SDK v2.0. However, thanks to the modularity of HemoKinect code, it can easily be adapted to new hardware that provides the 3D positions of the joints of interest as the above-mentioned sensors.
5. Conclusions
The role of exergaming in modern medical motor training and rehabilitation is overtaking traditional methods, as it allows remote patient supervision by exploiting the advantages of telemedicine. This work has validated the Kinect V2 for hemophilic patients’ exercise routines’ evaluation and tracking for the first time, using a completely newly developed software, HemoKinect.
HemoKinect relies on Microsoft’s Kinect SDK 2.0 to obtain 3D joint positions, which are used for joint angle calculation and COM estimation. These joint coordinates are mapped to the RGB image and streamed in real-time on a computer screen.
HemoKinect is able to successfully count: (a) elbow flexions and extensions; (b) knee flexions and extensions (squats); (c) step climbs and descents; and (d) measure balance performance towards eight different directions at three levels of difficulty. Additionally, a two-player mode is implemented. HemoKinect also allows saving and sharing the results (reports and graphs) via e-mail.
Although Kinect V2 is generally not aimed at medical application purposes where the measurement accuracy is paramount, its accuracy is enough to register hemophilic patients’ exercises and remotely track their progress and achievements to improve their physical fitness, which is a great step forward in telemedicine applied to hemophilia.
Future work will address the comparison of HemoKinect angle measurements for the joints of interest with other systems that are gold standards, such as clinical goniometers. Another possible line of work would be creating and evaluating more complex exercise routines for hemophilic patients.