From Reliable Sources of Big Data to Capturing Sport Performance by Ecophysical Variables
Representative Assessment Design and Technology
Technological advances provide the opportunity to further study and augment the understanding of the athletic performance as (i) they provide data that were not available in the past, and (ii) they allow computation of metrics that was not done in the past (see Chapter 1). For instance, machine and deep learning techniques allow overcoming the computation of‘low level’ indicators to add ‘high level’ indicators (e.g., intra- and inter-team organization), which can be done along the matches and/or the training sessions. In order to quickly provide feedbacks to practitioners about the behaviours and performance of the athletes, technological advances developed new devices and new applications/ algorithms (e.g., embedded sensors like smartwatches, smartphones, global positioning system - GPS, often complemented with accelerometers, constituting inertial measurement unit - IMU) and new tracking systems, notably based on progress in computer vision (e.g., multi-camera motion capture, TV broadcast tracking, deep learning approaches to 3D markerless motion capture, RGB depth camera). One issue with technological advances is that sensors always provide some data to scientists and practitioners, even when those data are poorly collected, poorly reflecting the core of practice or reflecting a decontex- tualized or non-representative designed protocol for collecting data, confirming the weakness of data-driven approaches (Couceiro et al., 2016). Moreover, a second issue is to have too much data, which request scientists to data reduction or dimensionality reduction by using techniques from AI (e.g., machine or deep learning). These issues raise the problem of how data reduction and data modelling do not reduce information meaning for practitioners. Moreover, when fuzzing data, reducing dimensionality, and modelling data, the outcome is sometimes complex to understand and to interpret for the coach, in the sense that the transfer from data analytics to sport intervention is far from direct. These risks could be overcome by appropriately designing procedures for data acquisition, defining the assessment task, the data collection method, and relevant technology. We argue that appropriately designed methodologies should root in the ecological dynamics framework. One theoretical pillar of this framework refers to perception-action coupling (Araujo & Davids 2015; Araujo, Davids, & Passos, 2007) and emphasizes the importance of a representative (experimental) task design (Brunswik, 1956).
Brunswik (1956) proposed the term representative design to advocate the study of psychological processes at the level of organism-environment relations. It means that contextual variables should be sampled from the organism’s environment so as to represent the environment where the organism behaves, and to which behaviour in the experiment is intended to be generalized. This definition of representative design emphasizes the need to ensure that learning, training, or experimental task constraints represent the task constraints of the performance or game/race environment that forms the specific focus of study. Representative design implies a strong emphasis on the specificity of the relations between the athlete and the environment, which could be neglected in data-driven approaches to behavioural sciences. The ability of performers to detect and use information from the environment to support their actions is predicated on an accurate and efficient relationship between perceptual and motor processes, referred to as perception-action coupling (Pinder, Davids, Renshaw, & Araiijo, 2011a). For example, Dicks, Button, and Davids (2010) compared movement and gaze behaviours ofsoccer goalkeepers in a typical video simulation with in situ research designs. The authors showed significant differences between task constraints that required verbal or simulated movements compared with those realized in situ, performing interceptive actions when facing penalty kicks. Such findings suggest that high representativeness of learning/training design would capture the perception-action coupling as it happens in competition and promote skill transfer between the learning task and the competition. Moreover, this study also points out the required caution when using video support for learning, training, and performance analysis, which if not properly used may decouple perception from action. In another example, Pinder, Davids, Renshaw, and Araiijo (2011b) emphasized the risk of using devices like ball projection machines, as they might remove remove key information sources from the performance environment (e.g., the movement of the bowler) and significantly affect the timing and control of interceptive actions in cricket batting. These authors observed significant differences between the practice task constraints, with earlier initiation of the backswing, front foot movement, downswing, and front foot placement when facing the bowler in contrast with facing the bowling machine. Therefore, more specific skill transfer may be expected when batting against a bowler bowling a ball (supported by the functional coupling between perception and action), while general skill transfer might occur when a bowling machine projects the ball. When using new technology for intervention and data collection, scientists and practitioners must be aware that this equipment might have an impact on the specificity of the skill transfer. Of course, when a bowling machine is used to project the ball, the reproducibility of the task is high and movement analysis can focus essentially on the batter. Conversely, when the bowler bowls the ball, movement analysis could also be directed to the bowler’s actions, as the batter picks up information from the bowler. For this to happen, the motion capture systems need to comprise both the bowler and the batter, which could lead to very heavy data collection (e.g., huge volume to capture and calibrate, many body markers to assess accurately upper body and arm). Moreover, when scientists decide to perform in-depth microscopic analysis, it sometimes demands heavy, intrusive (for the athletes), and time-consuming equipment for setting of the motion capture systems (e.g., multi-optokinetic camera system), which might disturb the data collection or even the athlete motion. Therefore, to guarantee representative experimental design, scientists should make a tradeoff among heavy motion capture systems, intrusive athlete equipment, and measurement accuracy. For instance, when using IMUs on limbs, segmental angles (i.e., angle between two limbs) are usually assessed, whereas the heavier motion systems that involve full-body markers allow building an anthropometric model to assess joint angles (Guignard, Rouard, Chollet, & Seifert, 2017b). This latter system is more accurate as it provides the exact centre of rotation of the joint but it is more intrusive for the athletes. Therefore, scientists and practitioners must always deal with the representativeness of the task design, the measurement accuracy, the limitations of the data collection setting, and the complexity and time consumption of the post-processing (i.e., data analysis) for providing feedback to the practitioners and stakeholders.
In short, to design a representative task for collecting data during the training session, practitioners need to design tasks that consider interacting constraints on movement behaviours (i.e., action fidelity/realism), adequately sample informational variables from the specific performance environments (i.e., relevant affordances), and ensure the functional coupling between perception and action processes, where progression towards task goal is evident (Araujo & Davids, 2015). Moreover, the technology for data collection and the computational methods for data analysis need to be meaningful for the practitioners. To enhance representative design, tasks should be viewed as: (i) complex and (ii) dynamic, to provide learners with opportunities to explore a variety of task solutions that evolve over time; (iii) novel and related to achievable goals; (iv) supportive of active perception; and (v) providing sufficient access to key sources of information in the surrounding environment (Davids, Araujo, Hristovski, Passos, & Chow, 2012). With these aspects in mind, the present chapter reviews the different design methodologies for data collection, defining the task, the procedures, and relevant technology, with emphasis on reliability (homogeneity or consistency between two systems; stability or reproducibility between test-retest, for instance, inter- and intra-observer equivalence). When reliability is considered, accuracy and precision could be assessed. Accuracy relates to the difference between the measurement and the part’s actual value, while precision describes the variation when the same part is measured repeatedly with the same device. Finally, we also look for the validity (degree to which the equipment or capture systems measure what they are theoretically expected to measure) of the different technological devices.