1. Introduction
Elbow and shoulder injuries among baseball players, in particular pitchers, continue to be a concern despite maximum pitch count recommendations and regulations [
1,
2,
3,
4,
5,
6]. Ligament and muscular damage at the elbow and shoulder has been associated with the repeated micro trauma sustained by these structures during the demands of high-speed throwing and pitching [
1,
7,
8]. In addition to financial costs associated with ligamentous injuries, typically requiring surgical treatment, functional day-to-day limitations and a long rehabilitation process create further loss to an athlete and/or an athlete’s professional organization. Measurements of accelerations and angular velocities per segment, plus computed torques and forces on the joints during pitching, may lead to better development of injury avoidance and return to sport after injury programs. Currently, optical motion capture is the standard tool that sports medicine biomechanists and clinicians use to study the mechanics of motion and their correlation with injuries. These systems provide data to guide diagnosis, treatment, training modifications, return to sport, or removal from training.
New technology is advancing motion capture to wearable sensor systems. The quality of these systems ranges from gadgets with limited or no calibration to accurate scientific tools. Whether lab-based or body-worn, current technologies are limited by sampling body segment motion collection at rates too slow to fully capture the ballistic human motion performed during pitching. The act of pitching includes body segment motion which is relatively slow at the start of the activity, creating a base for transferring momentum through the body in a proximal-to-distal pattern out to the throwing arm [
9,
10]. Pitching also includes the fastest recorded human body segment movement; the arm segment rotating about the shoulder joint measuring over 7000°/s [
11]. In recent years, physicians and team managers have observed, with attention and enthusiasm, the possibilities brought by new technology and methods for quantifying and qualifying high-speed sport performance [
12,
13,
14]. Using systems that evolved from our initial prototypes fielded in 2006, this paper reports one of the earliest efforts, to our knowledge, of providing reliable sports data using portable wireless wearable electronics that leverage an ultra-wide-range wireless multipoint inertial and magnetic sensor array.
As introduced above, most quantitative athletic biomechanical analyses still rely on manual video inspection or commercial marker-based optical systems [
15,
16], which consist of near-infrared camera arrays measuring at up to hundreds of frames per second, compromising resolution for capture speed. Setting up and calibrating a camera-based tracking system is time consuming, and the stability of the data processing can be affected by visual artifacts, occlusion, and changing background light. The expenses to purchase, maintain, and operate camera-based laboratories also limit access to biomechanical pitch analyses for many institutions and athletes. The accepted standard motion capture systems are mostly indoor lab-based equipment setups, which limit simulation of the outdoor game environment and potentially the athlete’s performance. Commodity depth-sensing cameras, as embodied in the Microsoft Kinect™, have had some application in sports analysis [
17], but range, speed, and accuracy limitations have constrained their capability. Active magnetic trackers are light-insensitive, but susceptible to distortion from conductive and/or ferrous metal and present very limited range of operation, as well as often inclusive of tethered cabled sensors [
18]. Mechanical measurement methods, such as goniometers [
19] and exoskeletons [
20], require the body to be restrictively cabled up or constrained. Vests, shirts, and garments, generally wired with embedded inertial, bioelectric, and fabric sensors, have likewise been explored and adapted for motion capture, including athletic sensing [
21,
22,
23,
24,
25]. One example is a system composed of a single inertial sensor applied onto the elbow via a skin-tight sleeve [
26,
27] for athletic applications. Also, high-quality flexible goniometers with embedded inertial units have recently become available [
28].
Sensors of nearly all types have grown smaller and cheaper, enabling their seamless integration into nearly everything, as envisioned decades ago by the pioneers of Ubiquitous Computing [
29]. Inertial systems have a limited history in basic motion and biomechanical research, dating back to the 1970s [
30], before integrated miniature accelerometers were available. Wired and wireless wearable inertial systems have appeared commercially over the last decade (e.g., [
24,
25,
26,
28,
31]) and in research (e.g., [
32,
33,
34,
35]), but have been mainly applied to non-ballistic motion capture, where the average motion speed is typical of human gait, as opposed to high-intensity sports analysis, only very recently providing the ability to capture high speed motion [
36]. Some researchers use inertial technology to only recognize posture and activity, dispensing the need for high range sensors and joint angle computation [
37], albeit at an information sacrifice. Additional information can be found in these recent review articles discussing the use of inertial sensors for lower limb movement [
38,
39], generic human motion [
40], and sports [
41].
However, while the product market has been successful in putting these small wearable devices on athletes and moving the athlete out of the lab setting, the data application in sport is still constrained by range and sampling rate [
42]. To address the challenge of quantifying the high-speed stresses incurred on the upper extremity during throwing, specifically baseball pitching, we set out to create a new inertial measurement unit (IMU) that can capture 3D motions occurring at both low and high speeds. Accordingly, we have developed a wearable inertial sensor platform to enable end-to-end investigation of high-level, very wide dynamic-range biomechanical parameters. Unlike commercially available wireless systems that have been designed for motion capture, our device has extremely high dynamic range and exhibits precise synchronization across multiple wearable nodes.
Using the state-of-the-art camera-based motion capture systems, shoulder and elbow distraction forces are calculated using the second derivative of the measurement system data, i.e., linear acceleration, and inverse kinematics. Unfortunately, the derivatives of orders greater or equal to two have high levels of noise, often resulting in limited or no physical significance, unless the original data—in position units—is filtered down to 10–20 Hz. This filtering damps rapid signal variations, and hampers proper inference of higher-order derivatives that happen during excessive joint load. Accordingly, we assert that classical optical systems are limited in producing meaningful assessment of these forces.
Finally, we introduce the concept of jerk to the evaluation of pitching mechanics using our IMU system. As defined in classical mechanics literature (e.g., [
43]), jerk is the third derivative of position, and it expresses the rate of change of acceleration (as opposed to acceleration itself which is the rate of change in velocity over time). We suggest that the rate of change of acceleration may be more related to microtrauma than the absolute value of acceleration, which is canonically used to obtain force metrics. Given the assertion about the noise inherent in the second derivative of the positional data to calculate acceleration, calculating a third derivative of positional data has been effectively prohibited in previous optical-based biomechanical evaluations of pitching. We hypothesize that meaningful jerk data could be obtained from our multi-segment inertial system.
In this paper, we present the scientific requirements needed and the steps taken to build a robust and accurate wearable sensing system with high autonomy and portability for baseball and provide initial comparisons to an optical motion analysis system. In baseball, pitch type is often distinguished based on the grip of the baseball and the motion of the hand and forearm. Wrist flexion and extension rely on the action of the larger muscles in the forearm, some of which cross the elbow joint. Therefore, we included wrist joint force and hand angular velocity in our analyses. Elbow valgus/varus torque, and shoulder and elbow distraction forces were biomechanical metrics selected for comparison based on their established connection to shoulder injuries and UCL (Ulnar Collateral Ligament) sprain [
44,
45]. A series of studies were performed to address the following aims: (1) Compare the raw output of wrist force, wrist angular velocity, shoulder angular velocity, and shoulder and elbow distraction forces between an optical marker-based system versus our developed inertial system. (2) Investigate the influence of filter processing on optical system data compared to the data from the multimodal wide-range IMUs. (3) Investigate the feasibility of using shoulder jerk as a metric from the IMU wearable system for identifying differences in stress at the shoulder joint across pitch types.
4. Discussion
Current optical systems have allowed clinicians to gain insight into the velocities, forces, and torques placed on joints susceptible to injury during ballistic motions and repetitive microtrauma. These estimations, however, are limited by factors such as data filtering methods, lower camera resolution during peak speed of optical camera systems, and artifact motion due to soft tissue movement caused by the high motion speed. Our data is the first to compare a multimodal wearable IMU system to that of an optical system during the ballistic action of the overhand pitch and to investigate the influence of filtering of optical systems compared to IMU data.
Smoothing positional data allows for the use of the optical data for computing kinetic quantities [
10,
45,
57,
58,
59,
60,
61]. However, qualitative comparisons of filtered pitching data show loss of potentially meaningful information with the use of standard filters. When looking specifically at valgus elbow torque, our limited dataset demonstrates that calculated optical system elbow torques fall short of those calculated from the inertial system and suggests that stresses on the elbow may be higher than previously evaluated. Some of this limited resolution is related to the capture rate of the systems and the more direct relationship between inertial measurements and forces and torques. The higher capture rate of our sensor-based system (1000 Hz) versus optical systems that commonly capture pitch trials between 240–500 Hz, appears to allow for a better description of the peak dynamics. These results raise an important concern for the field of biomechanics. It is common practice to low-pass filter the optical data to smooth out inherent jitter in the reconstructed optical data. However, the conservative cutoffs used on these filters appear detrimental to the data output, limiting the ability to capture peak dynamics in high-speed athletic motion.
This study also introduces the concept of ‘jerk’ to the evaluation of pitching biomechanics. Jerk, the change in acceleration over time (the 3rd derivative of positional data) cannot be properly derived from optical data of pitching due to the noise within the data, but as shown in this paper, can be usably calculated from IMU data. Measured levels and timing of jerk may offer new assessments of soft tissue injury risk. Evaluation of jerk forces at the shoulder in this study showed a trend toward high jerk forces with sliders compared to fastballs [
46], but this did not reach statistical significance with our limited sample of pitchers. A deeper insight into the implications of jerk forces in the shoulder and elbow will benefit from more data collected from such high-rate, wide-range wearable inertial systems.
Clinical speculation is that ligament, tendon, and muscle tissue integrity breaks down due to repetitive microtrauma from repeated high speed pitch deliveries [
5,
64]. Our preliminary studies have seen indications of frequent changes in acceleration occurring throughout the pitching motion [
46]. Although some of this may arise from soft tissue artifacts, such frequent changes in node-measured peak angular velocities indicate accelerations and decelerations not observed in the optical motion data, and may directly contribute to these microtrauma tissue stresses. If this is the case, then the true severity of the trauma is under-measured with use of optical systems. High-speed, wide-range inertial data has the potential to provide both clinician and athlete accurate force/torque information, and in combination with clinical measures/symptoms, offers a unique way to monitor joint stress and joint health.
Although our studies posit a strong argument for the superior veracity of wearable inertial sensors over optical tracking systems for accelerations, forces, torques and parameters inferred from them, one must admit that neither of these systems provide the absolute ground truth. Both systems, camera-based and wearable sensors, measure physical quantities at the skin overlay, and from this data, we are estimating quantities for the underlying rigid bodies, i.e., bones, such as forces and torques. Therefore, there is no noninvasive method today that is able to provide direct, dynamic information of the hidden rigid body structures, especially for fast motion like pitching.
Hence, like all motion analysis studies, this study is limited by soft tissue artifacts (STAs) [
46]. For baseball pitching, the ballistic nature of the motion magnifies the effect of STAs. In optical motion-capture systems, STAs are considered the most troublesome source of error [
65]. Soft-tissue artifacts manifest in several different ways, such as the inertial reaction of the sensor against elastic skin, and rocking of the sensor as muscles move, contract/extend, and deform underneath during an extreme athletic gesture like pitching [
66]. Previous work has shown that upper arm axial rotation (humerus internal-external rotation), is the upper arm motion most affected by STAs [
67]. Some studies proposed the mechanical coupling of forearm and upper-arm to compensate for upper-arm artifacts [
68]. This technique mitigates the noise in the upper arm data. Even if successful compensation for this coupling is obtained, the solution is limited to motion where the elbow does not reach full extension [
68]. This argues that single segment IMU evaluations may not be as accurate as multi-segment models in the future [
69].
Wearable sensors offer the ability to collect data across body segments and discover motion patterns that may correlate fatigue with risk of injury to the elbow or shoulder [
10]. The inter-segment timing sequence data provided by this system is more accurate than optical systems by nature of the IMU’s high frequency response and rapid synchronized sampling rate. The sensors hence nicely yield the relationship between the timing of each segment’s peak angular velocity, as each transfers its momentum out to the hand during both pitching and bat swinging [
46].
Magnetometer-augmented IMU-derived position can be performed when using this system on its own and ‘in the field’, but is not as accurate as position obtained from optical motion capture systems. However, we have seen that the sensors used in the present study provide superior measures of angular velocity, acceleration, and jerk. If one wants to obtain a best set of kinematic and kinetic data, it seems likely that the optimal approach would be to fuse the position and orientation data from a motion capture system with the inertial data obtained from the sensors. One method for fusing the inertial and motion capture data would be to use well-established estimation algorithms based on Bayesian inference that provide a principled way for making optimal inferences from the inertial data and the motion capture system [
70]. The use of probabilistic sensor fusion as outlined by Todorov [
71] may improve the quality of the joint force and moment estimates, as well as provide the dynamics consistency required for the development of the musculoskeletal models needed to estimate the muscle and ligament forces during a baseball pitch [
72].
Using the state-of-the-art camera-based motion capture systems has limitations in data capture rate. The shoulder and elbow distraction torques and forces from such systems are calculated using the second derivative of the measurement system data, i.e., linear acceleration, and inverse kinematics. Unfortunately, the derivatives of orders greater or equal to two have high levels of noise, often resulting in limited or no physical significance, unless the original data—in position units—is filtered down to 10–20 Hz [
46]. This filtering damps rapid signal variations, and hampers proper inference of higher-order derivatives that happen during excessive joint load. Accordingly, we assert that classical optical systems do not permit complete insight to the ballistic motion of the baseball pitch, and we hypothesize that a multi-segment inertial system can produce these quantities with increased precision over optical systems.
Our nodes seemed to be mechanically adequate for evaluative scenarios. Skilled motor performers are known to be able to adapt to new setups and interferers around their body easily. The nodes are light and the players did not express concerns related to movement constraints. Most players were queried and there was not a single report of the IMU-based system hindering their performance. The common response indicated that after a few pitches the player “got used to it” and “did not feel it at all”. During our experiments, the players completed their typical bullpen session without complaints for their routine of approximately 50 pitches. We did not design the mechanics of our nodes to be used in actual competitive games, which involve deeper physical and regulatory constraints.
The underlying hardware for wearable IMUs is steadily evolving. The now-common integration of all inertial components onto a single die has been driven by the large mobile devices market, and although this enables a much more compact form factor, such combined devices do not yet provide the extreme dynamic range we need here. The eventual development of log-scale accelerometers and gyros will enable ultra-wide-range, high-resolution devices that do not need redundant measurements at different scales, and will be well suited to measuring athletic gesture in high-intensity sports. Finally, the continual evolution of stretchable electronics is enabling devices such as ours to be embedded in a conformable form factor better suited to mounting on the body [
73,
74,
75], with the caveat that STAs and inertial reaction in intense athletic motion may introduce more effects in a deformable platform.