A Knee Rehabilitation Exercises Dataset for Postural Assessment using Wearable Devices

Exploratory data analysis

The objective of this section is to validate the technical quality of the built dataset by performing an Exploratory Data Analysis (EDA). To this end, the extracted statistical properties (i.e., features) should have a bio-mechanical explanation, while taking into consideration that all the participants suffered from a knee pathology (i.e., not healthy). Thus, some variations when comparing the muscle activations or kinematics between the healthy and injured leg, even in the correctly performed exercises, are expected. The main characteristics of the performed EDA are summarized using statistical graphics in Fig. 7, where the upper part (a) includes the variations of squat, the middle part (b) the variations of seated leg extension and the lower part (c) the gait variations.

EDA performed on the squat variations (a), seated leg extension variations (b) and the gait variations (c).

In Fig. 7a at the left the standard deviation σ of the rectus femoris activation is depicted, where the difference between the healthy and the injured leg for the Squat_WT is much higher compared to the other squat variation, since the user’s weight is transferred on the healthy leg. Similarly, the σ of the hamstring activation is higher for the injured leg when performing Squat_FL. Finally, at the right of Fig. 7a we can see the estimated Euclidean distances between the obtained accelerations of the Z-axis (forward view of the user, see Fig. 1b) of the healthy and the injured legs for every sensor placement. The most significant difference is observed for the case of the sensor placed on the gastrocnemius, where the mean Euclidean distance estimated for the case of the Squat_FL variation is more than x2 larger than that of the Squat and more than x1.5 larger compared to the corresponding value of the Squat_WT. This is expected if we consider the fact that this body part along with the tibialis anterior, where the same pattern is also observed, is extended more in the Z-axis compared to the rectus femoris and hamstrings.

Moving to Fig. 7b at the left we observe the estimated Euclidean distances between the obtained sEMG signals of the healthy and the injured legs per sensor placement. For all the placements, the distances for the Extension_NF variation are smaller compared to the other two variations, meaning that, since the muscle activations of the healthy leg should be close to zero, the patient does not fully extend his/her leg leading to smaller activations. In the middle, the maximum tibialis anterior Y-axis accelerations are depicted for the case of the injured leg, where again the not fully extended leg leads to less movement of the tibialis towards the vertical axis. For the case of the Extension_LL variation the most significant statistical difference when compared to the correct extension is observed on the σ estimated for the X-axis acceleration (i.e., vertical side view) of the hamstring. This is expected from a biomechanical point of view since the patients in order to fully extend their legs, lifted their limb from the chair (Fig. 3b2), thus, even though the tibialis acceleration for these two exercises is almost identical, the performed leg abduction when executing the Extension_LL is captured by the X-axis accelerometer revealing the identity of the exercise.

At lower left part of the EDA illustration (Fig. 7c) the estimated σ of the tibialis sEMG signal is shown. As aforementioned, since the data are originated from people suffering from knee pathologies, even for the case of normal gait the σ values vary slightly between the injured and the healthy leg. The difference is more observable for the case of the Gait_HA, where the muscle activation of injured leg’s tibialis anterior is higher, since the patient keeps his foot flexed and consequently does not relax this muscle. Observing the plot at the lower middle, it is evident that the σ of the accelerations in all axes is much higher when the patient was walking normally, since in the wrong variations the patients tends to control its leg movement (flexion and/or extension) to avoid any potential pain. Finally, we estimated the kurtosis of the Y-axis acceleration of the gastrocnemius. Kurtosis generally indicates whether the distribution or the produced values is thin (high kurtosis values) or broad (small kurtosis values) and can be used as an indicator of normality, with the kurtosis of a normal distribution being equal to 3. Here, normal gait produces almost identical kurtosis values for the injured and the healthy leg, with the average value being almost close to 3. For the case of the Gait_NF variation the healthy leg’s kurtosis are higher than those of the injured, since during a gait cycle the patient tends to lower the injured leg and lift the healthy leg to support the not fully extended injured limb. The opposite pattern is shown by the Gait_HA, due to the abduction the patient performs on the injured limb (Fig. 4c3) leading to more sharp vertical movements (i.e., quick lift-offs).

Comparison with public wearable-based datasets for postural assessment

In the current section, we encounter postural assessment datasets that are only publicly available (i.e., we exclude those that are not^13,31). Although, wearable-based datasets including both IMUs and sEMG exit these are designed for human activity recognition^17,18,19, and/or for lower limb joint’s angle estimation^{11,16,17,18,32} (useful to lower-limb prostheses and exoskeletons). They are not designed for postural assessment and they use healthy subjects in the experiments. To the best of our knowledge, KneE-PAD is the only publicly available dataset that is designed for rehabilitation assessment while containing both modalities and data from patients having knee pathologies.

Starting with unimodal datasets, PHYTMO¹² includes the IMU recordings of 30 healthy volunteers performing a total amount of 6 rehabilitation exercises (upper and lower limb) and 3 gait variations. Each participant performed two sessions with a minimum of 8 repetitions in each one, producing 7,076 files in total. UI-PRMD¹⁰ is a vision-based dataset consisting of 10 rehabilitation movements, executed by 10 healthy individuals. Each movement was repeated 10 times (correct or incorrect) in front of two sensory systems for motion capturing an optical tracker, and an RGB-D camera, leading to 2,000 files. Thus, the data provides along with the exercise type, the body joints’ positions and angles. A completely different approach exploiting a millimeter-wave (mmWave) radar and an RGB-D camera as a reference (estimated joints’ angles and positions) is introduced by MARS³³. The dataset contains the mmWave radar-produced 3D point cloud (including also the Doppler velocity and intensity) obtained by 4 healthy subjects performing 10 specific rehabilitation movements, while being supervised (no incorrect movements included).

When it comes to multi-modal dataset, FineRehab²⁰ comprises 16 exercises executed by 50 participants. It is worth mentioning that this dataset contains both healthy individuals and patients with musculoskeletal disorders. The total dataset consists of 4,215 files captured by two Kinect (RGB-D) cameras and 17 IMUs, offering also the RGB-D based estimations of joints’ angles and positions. Similarly, Palermo et al.³⁴ exploited the same sensor modalities to develop a dataset consisted of 14 healthy participants walking with a wheeled robotic walker. Consequently, this dataset is suitable for gait analysis, containing the 3D joints’ angles, since other rehabilitation exercises are not included.

Limitations

After conducting the literature review presented in the previous subsection, we concluded that even though the collected dataset offers the advantage of including multi-modal patients data, it includes less sensor modalities compared to others^{11,16– 18}. In particular, vision markers could be very useful in order to capture the ground truth joints’ angles of the patients, making the dataset also applicable for the development of 3D pose estimation algorithms. In addition to this, upper limb rehabilitation exercises are not included, reducing its application to knee rehabilitation exercises.

Finally, there is a statistical bias on the patients’ data displayed vividly in Fig. 8. KneE-PAD is not highly imbalanced in terms of the participants gender, however, it is biased when it comes to the participant’s age and knee pathology. In particular, the included male patients are significantly younger than the female ones, while suffering a ligament injury, not osteoarthritis that female patients tend to do. Even though, this statistical bias is expected³⁵, it limits the machine learning tasks the dataset could be used for. For example, for the case of knee pathology identification, although the gender and height could be left out of the feature set, the latter could be revealed by the range of patient movement of the sensor’s placement.

Patients’ demographics in terms of gender (blue or orange color), height (Y-axis), age (X-axis), pathology (“x” marker for osteoarthritis and “o” for ligament injuries), and weight (size of the marker).