Abstracts Track 2024

Area 1 - Computer Systems in Sports
Area 2 - Health and Support Technology
Area 3 - Sport Performance and Support Technology

Area 1 - Computer Systems in Sports

Nr: 10
Title:

Kinematic Analysis of Golf Swings: A Comparative Study of Multiple Three-Dimensional Human Pose Estimation Models Using the MMPose Framework
Authors:

Cavan Aulton, Ben W. Strafford, Keith Davids and Chuang-Yuan Chiu
Abstract: Traditional approaches to full-swing analysis in golf typically uses marker-based motion capture systems which are expensive, include time-consuming set-up procedures and require golfers to perform in a laboratory environment (van der Kruk & Reijne, 2018). Contrastingly, three Dimensional (3D) Human Pose Estimation (HPE) is a field method of predicting the position and orientation of body parts from images and videos (Zheng et al., 2022). The MMPose framework offers pre-trained models and enables users to implement different HPE models effectively (Grover et al., 2022). MotionBert (MB) performs well in capturing complex motion and handling occlusions because of its ability to consider multiple frames simultaneously (Zhu et al., 2023). VideoPose3D (VP) processes video sequences which improves prediction smoothness and is an efficient model which is beneficial for applications which require near real-time processing (Wang et al., 2021). Due to varying model architectures and potential similarities in application outcomes, understanding the differences in approaches to inform future practice is crucial. Therefore, this study aims to compare the ability of MB and VP models in conducting kinematic analysis of the full swing in golf. To do this 143 full swing clips were trimmed from the GolfDB dataset and analysed with both models using the MMPose framework. Video clips are treated as input data, feature extraction is first conducted on 2D images, then 2D pose estimation using the RTMpose model was conducted on these features, then a lifter (MB or VP) is used to predict the 3D joint locations. The 3D joint locations were then used to extract angular variables frame-by-frame, then angular displacements were calculated for kinematic variables correlated with golfing performance: the left and right arm to trunk, left and right elbow, left and right wrist, left shoulder adduction in the transverse plane, pelvic girdle rotation, shoulder girdle rotation, and axial trunk in the transverse plane (Horan et al., 2010; Parker et al., 2022; Zheng et al., 2008). Overall, the results for all angular displacement values were similar between the models with the largest difference being between the angular displacement of left elbow MB 113.8° ± 26.1° and VP 85.0± 20.9°, t(144)=18.326, p = 0.000, full results can be found in table 1 of the appendix (in supplementary material). The inter-class correlation coefficient ranged from moderate to excellent (0.363 – 0.885), demonstrating a mixed agreement in assessing joint movements between the models, full inter-class correlation results can be found in table 2 in the appendix (in supplementary material). These results indicate a high level of consistency between the models, but face problems providing consistent results for axial trunk rotation in the transverse plane and left arm to trunk angular displacement. The reliability differences may occur from occlusions caused when the left arm crosses over the trunk, complicating model differentiation between the arm and trunk. Both models are trained on large datasets containing a variety of sporting and non-sporting scenarios, but the amount of golf specific images is limited. Future research should validate these approaches by comparing them against marker-based motion capture to determine which model provides the most accurate variables, whilst adapting models to track the club to provide more performance metrics to practitioners.
Nr: 25
Title:

Deep Reinforcement Learning for Muscle Strength Improvements
Authors:

Lois Wakili, David Rogerson, Stephen Thompson and Chuang-Yuan Chiu
Abstract: Introduction Resistance training (RT) benefits sports performance and health by enhancing muscle mass and strength. Optimal adaptations require the precise manipulation of RT variables (frequency, intensity, time, etc.) through carefully crafted programmes designed specifically to benefit the individuals participating in them. Individualised programming necessitates specialist knowledge, limiting access to the general public. Hence, there is a need to explore advanced techniques such as reinforcement learning (RL). In RL, a computer programme (an agent) interacts with its environment and learns to make better decisions by receiving rewards or penalties based on its actions. Recent advancements in RL with deep learning models have demonstrated the potential for personalised healthcare recommendations. This technique, however, has yet to be used to deliver personal RT intervention programmes. Thus, this study aims to explore the potential of RL in prescribing RT interventions. Methods The dataset for this study was extracted from the Fit20 dataset, containing anonymised data from 14,690 individuals over six years. Participants performed weekly strength training exercises on three machines (leg press, chest press, and lateral pulldown), recording weights and exercise durations. The agent; a Deep Deterministic Policy Gradient (DDPG) algorithm, was initially trained on data from one individual, comprising 313 sessions, to identify the potential of using RL for personal RT intervention. States were defined as the weights on each machine, and actions were the time spent on each machine. A reward function was defined by the differences in weights lifted after four consecutive sessions. Results The resulting model provided metrics such as the cumulative rewards (ranging from 7.99 to 17.13) and corresponding state-action pairs with the maximum reward (weights in kg [68.0, 27.5, 35.0] and duration in seconds [199.9, 51.98, 192.50]), thus indicating the training programme where this particular individual saw the most strength improvements. Discussion This study demonstrates the feasibility of using reinforcement learning to optimise strength training by focusing on specific combinations of training variables that improve long-term strength outcomes. The dataset used lacked variability because only one RT variable represented the actions, and there was little anthropometric information about the participants. As a result, additional research would be required to collect comprehensive baseline data from more individuals, with frequent measurements taken during training programmes. This would allow for the development of a more robust and flexible model capable of providing ideal training regimens for improved muscle mass and strength.
Nr: 53
Title:

Lateral Ankle Sprain: Product of a Functional Network System with Architecture - New Prevention Method and Device
Authors:

Daniel Molitorisz
Abstract: OVERWIEV Lateral Ankle Sprain (LAS) injuries have long been regarded as the accidental result of a series of unfortunate coincidences. Either one or another factor was considered to be the key responsible element for occurrence of the injury. Attempts were made to invent effective preventive techniques based on these theories, however they proved to be unsuccessful.  We've discovered an universal trigger event to lateral ankle sprain (LAS) which isn’t sports specific, but detectable prelinked to every LAS injury. Identified and separated we found that lifting of the hallux (LoH) while loading response-, midstance phase of the step is an obligate predictive factor to LAS. Our theory and observations in practice presentes LoH as product of a network system which function we can effectively block. The architecture of this network corresponds to the laws of the growing scale-free real network model described by Albert-László Barabási. (BA-model) This novel finding can be seen as method to avoid the most common musculoskeletal injury. METHODS Participants performed bodyweight double leg calf raise, normal gait and run presenting loading response phase of step in various situations. 550 participants tasks (age: 10-43y) were videorecored and analyzed. LoH and contact force alterations under hallux were recorded. We collected and analyzed the videos from LAS on the internet RESULTS The occurence of losing balance is 73 times higher if hallux dosen’t support the loading response phase (χ²(1) = 186,134, p<0,0001, OR:73,119) presenting unstbale (equilibrium) pre-injury-posture. Foot inversion-adduction were detected, heel tilted while lowering contact force under hallux. CONCLUSIONS: The summary of linked occurrences creates the pre-injury complex which produces the pre-injury-posture. The pre-injury-posture: LoH is an unstable equilibrium point. Used the functional network model of Barabasi Albert, LoH represents the result of linked nodes built up system. The injury itself is triggered when tilting out from LoH. When blocking the main occurrence of pre-injury-posture, we can effectively avoid lateral type ankle sprains. Detecting the steps when contact force lowering under the big toe is a predictive factor suspicious to LoH, therefor to lateral ankle sprain. We developed and designed an intelligent insole whit bluetooth connection to smartphone and watch capable notice and sign suspicious steps to lateral type ankle injury.
Nr: 69
Title:

Detection of Hypoglycemia in Elite Endurance Athletes Using CGM; Impact on Sleep, Training and Recovery
Authors:

Filip J. Larsen and Lina Spetz
Abstract: Continuous glucose monitoring (CGM) has gained significant attention in the sports community, particularly among endurance athletes who use skin-mounted sensors for real-time glucose tracking. However, despite its growing use, a clear framework for interpreting CGM data in the context of athletic performance and health remains underdeveloped. There is limited understanding of the physiological impact and practical implications of blood glucose fluctuations in athletes. Emerging research indicates that glucose regulation in endurance athletes differs from that of non-athletic populations. Additionally, the metabolic demands of endurance training, including variations in energy and carbohydrate availability, can significantly affect glucose regulation, complicating the interpretation of CGM data. Methods: In this study, we followed 19 elite endurance athletes over the course of one year, tracking their training with GPS watches, sleep with Oura rings, blood glucose levels with CGMs, and subjective recovery ratings through a mobile app questionnaire. Results: The cohort exhibited an average 24-hour blood glucose concentration of 5.28 ± 0.22 mM but spent, on average, 1 hour 44 minutes (range: 7 minutes – 2 hours 54 minutes) per day in the hypoglycemic range (<4 mM). Additionally, they spent an average of 28 minutes (range: 7 minutes – 51 minutes) per day in the hyperglycemic range (>8 mM). Athletes who experienced more time in hypoglycemia reported lower subjective sleep efficiency, reduced REM sleep, and increased deep sleep, despite no differences in total sleep time or bedtime compared to those with less hypoglycemia. Interestingly, subjective recovery scores and total training load were not significantly different between athletes with higher or lower hypoglycemia exposure. Self-reported food intake also did not differ between these groups. At the individual level, longitudinal analysis revealed that nights with increased (>2h) compared to less (<0.5h) hypoglycemia were associated with mild sleep disturbances; a 0.7% lower sleep efficiency, and 7 minutes more awake time per night. The subjective training load was significantly higher on the day preceding the night with more hypoglycemia. Discussion: The occurrence of hypoglycemia appears highly variable among athletes, with some individuals being more prone to hypoglycemic episodes, particularly during sleep between 3 AM and 6 AM. Nighttime is a critical period for recovery, and disturbances in sleep can negatively affect the body's restorative processes. Identifying athletes more susceptible to hypoglycemia is crucial, and CGMs may assist in early detection and personalized interventions.

Area 2 - Health and Support Technology

Nr: 48
Title:

Estimation of Human Somatotype Components Based on Depth Camera Data
Authors:

Krzysztof Przednowek, Élvio Rúbio Gouveia, Maciej Sliz and Tomasz Krzeszowski
Abstract: Background: Anthropometric parameters are measured using several specialized devices and measuring instruments, which requires properly trained personnel and is time-consuming. Given the wide applications and drawbacks of the methods currently used in anthropometric measurements, it becomes reasonable to develop methods that would allow automatic and rapid measurement of many parameters and thus automatic determination of somatotype. Such a solution could be the acquisition of 3D scans of the human body using a depth camera, and then determining the mentioned parameters from the obtained 3D scans. Aim: The aim of this research is to develop methods for estimating body composition components from depth camera data. The depth camera data represent scans of the human body, from which the circumferences and areas of each segment are calculated. Material: The material of the study consisted of 129 patterns (3D scans) obtained from men aged 18-28. After recording the 3D scans, segmentation of the 3D scans of the human body is carried out, during which segments corresponding to the following body parts are separated: arms, torso, thighs, and calves. Circumferences and areas are calculated for the designated segments, which form the basis for determining models estimating the various components of somatotype. Methods: The study used the classification of somatotypes according to the Heath-Carter method. This classification provides for three types of somatotypes: endomorphy, mesomorphy, and ectomorphy. Somatotype is defined as the values of the three components representing each of the somatotype types. Classical measurements of girth and folds provided the benchmark for the regression models. The following regression models were used to estimate each somatotype component: classical linear regression (OLS), Ridge regression (Ridge), and LASSO regression. All models were optimized for structure parameters. RMSE error calculated by leave-one-out cross-validation (LOOCV) was used as a quality criterion. All calculations were performed in the GNU R environment. Results: The study showed that for each somatotype component, the lowest error is generated by the LASSO regression model. The endomorphy component is estimated with an error of 0.709, the mesomorphy component 1.196, while the ectomorphy component is estimated with an error of 0.450. Conclusions: The obtained errors were characterized by relatively small values. Concerning the group averages, the individual errors were: 23.7% for endomorphy, 15.5% for mesomorphy, and 20.2% for ectomorphy. Based on the results obtained, it can be deduced that the most difficult task is the estimation of the mesomorphy component. However, it should be remembered that the accuracy of the models is largely related to the accuracy of the 3D scan. Increasing the accuracy of estimation will be closely related to the expansion of the pattern base so that the representation of individual somatotypes is comparable.
Nr: 65
Title:

Stride Shortening in Patients with Ankylosing Spondylitis
Authors:

Igor Gruic, Vedran Brnić and Frane Grubisic
Abstract: INTRODUCTION Key spatiotemporal quality of gait is adequate stride length. Bilateral symmetries in kinetic and kinematic parameters are not prerequisites for healthy gait - if considered trough absolute values. However, tendency towards symmetries in all medical and sportive phenomena and expressions - are dominant determination wherever possible. Although anatomy of internal organs and structures reveal helical and spiral functionalities of and within the body (helical hearth, unilateral dominance of head/leg etc), firstly - those features are insufficiently utilised, or secondly – disruptions of inherited asymmetries are clinically or scientifically not adequately measured and recognised by health professionals even if those represent probable quantified messenger of disease, in patients with e.g. ankylosing spondylitis. Aim of the study was to determine sources of stride asymmetries in patients with ankylosing spondylitis, with and without relation to healthy stride. METHODS The sample of subjects were 24 individuals - 12 patients suffering from ankylosing spondylitis (AS) disease, and 12 healthy persons (K). Average body height (Mean+/-SD) was 175.91+/-7.87 cm, average body mass 82.51+/-14.83 kg, with the average age of 43.04+/-9.12 years. Basic and derived anthropological characteristics of respondents were measured (feet length/width etc.), kinematic (stride length, angles in ankles, knees and hips) and kinetic pedobarographic (foot contact with derived forefoot, midfoot, and heel forces) parameters. Descriptive statistics (mean, standard deviation, etc.), were calculated, followed by normality test - Kolmogorov-Smirnov. The t-test for dependent samples was used to determine differences. All analysis were performed in the Statistica 14.0.0. tool. RESULTS Mean and standard deviation (Mean+/-SD) in variable stride length was 1.13. +/- 0.12 for AS group, compared to 1.27 +/- 0.12 in K group. The t-test for dependent samples between groups in variable stride length is t=-2.91; df=22; p=0.008. All angles in ankles, knees and hips were predominantly similar for both groups with regard to left leg, and slightly lower for right leg. Within kinetic pedobarographic parameters, differences among maximal forces were not statistically significant, with similar relative values, which was followed by same findings for foot contact with forefoot and heel, but not midfoot forces (t=2.19, df=22, p=0.038). DISCUSSION AND CONCLUSIONS Stride differences between two groups were expected outcome. However, underlying contribution for manifested asymmetries were not. Helical and spiral functional patterns - from anatomy to movement -were very straightforwardly represented in results. Although there were no statistically significant differences between angles in joints of left and right leg, there was a pattern of smaller angles and amplitude of movement of right leg, followed by non-proportionally greater relative forces of right mid-foot of AS compared to K group (suggesting obstruction of foot arcs). ACKNOWLEDGEMENTS Within Project CroRIS ID: 4807, previously approved by Committee for Scientific Research and Ethics of the University of Zagreb, Faculty of Kinesiology, Croatia, Date 28.09.2020, No:98/2020, and by Ethical Committee of Clinical Hospital Centre Sisers of Mercy, Zagreb, Croatia No:EP/19438/19-17. Date:5.12.2019.
Nr: 67
Title:

The Center of Mass and Center of Pressure Reflect Distinct Qualities of Postural Stability Under Fatigue and Non-Fatigue Conditions in Sedentary Employees
Authors:

Banafsheh Amiri and Erika Zemková
Abstract: Abstract Background: The center of pressure (CoP) is commonly measured using force plates to assess postural stability. However, this parameter provides limited information regarding the movement of the center of mass (CoM) in the lumbar region. Since recent advances in sensor technology allow the registration of CoM, we were interested in the relationship between these two parameters under different conditions. This study investigates the relationship between CoP and CoM variables measured while standing on stable and unstable surfaces under non-fatigued and fatigued conditions, assessing postural stability in sedentary employees. Method: Seventy-two sedentary adults (44 women and 28 men, aged 36.2 ± 8.1 years) were assessed using the International Physical Activity Questionnaire. Participants performed a modified Abt's fatigue protocol consisting of eight exercises targeting the core muscles. Before and after the fatigue protocol, CoP velocity and CoP displacement were registered using a force plate-based posturography system. Simultaneously, CoM velocity and CoM displacement were recorded using a Gyko inertial sensor system located on the trunk. Data were collected during tandem stance with eyes open, either on a force plate or on a foam mat positioned on top of it. Results: As expected, after core muscle fatigue, CoM velocity increased more than CoP velocity under both stable (22.07% and 18.45% respectively) and unstable conditions (25.16% and 15.05% respectively). Similarly, there was a greater increase in CoM than CoP displacement under both stable (31.33% and 17.23% respectively) and unstable conditions (29.52% and 21.03% respectively). Under fatigue conditions, CoP and CoM velocity were moderately correlated while standing on a stable and unstable support surface (0.30 and 0.59 respectively), resulting in 9.00% and 34.81% common variances, respectively. Likewise, weak to moderate correlations were found between CoM and CoP displacement when standing on a stable and an unstable surface (0.36 and 0.44 respectively), yielding common variances of 12.96% and 19.36%. Under non-fatigue conditions, there were moderate correlations between CoP and CoM velocity (0.32 and 0.65 respectively) leading to common variances of 10.24% and 42.25%. Furthermore, CoP and CoM displacement demonstrated strong correlations (0.62 and 0.62), both showing a common variance of 38.44%. Conclusion: The weak to moderate common variance between the CoP and CoM parameters indicates limited overlap under both fatigue and non-fatigue conditions. This implies that they may capture distinct aspects of postural stability. Utilizing initial sensors alongside the posturography system provides deeper insights into sway velocity and displacement in sedentary employees. Trial registration number: IRCT20221126056606N1. Date of registration: December 25, 2022. This work was funded by the UK Grant of Comenius University in Bratislava (Project No. UK/3155/2024, awarded to Banafsheh Amiri) and supported by the Scientific Grant Agency of the Ministry of Education, Science, Research, and Sport of the Slovak Republic, and the Slovak Academy of Sciences (Grant No. 1/0725/23).
Nr: 72
Title:

Post-Exercise Circulatory Occlusion Differentially Affects Soleus H- and V-wave Excitability
Authors:

André Gonçalves, Andréia Terra, Rodrigo Marques, Pedro Pezarat-Correia, Carolina Vila-Chã and Gonçalo Mendonça
Abstract: Introduction: In exercising humans, fatigue is well known to be associated with the activation of type III/IV muscle afferents. However, whether their role on fatigue results from inhibitory modulation at the spinal or supraspinal level, has not yet been fully understood. Post-exercise circulatory occlusion (PECO) allows to investigate the influence of maintained metabolic feedback (i.e., originating form type IV muscle afferents) on central nervous system failure site of action. In the lower-leg, both soleus H-reflex and V-wave have been shown to decrease after fatiguing sustained submaximal contraction. Despite this, it remains unknown if continued muscle afferents-induced discharge causes the same modulatory effects at spinal and supraspinal level. Therefore, the purpose of this study was to explore the effects of PECO on soleus H- and V-wave excitability immediately after lower-leg exercise. Methods: Sixteen healthy males (age: 24.6 ± 3.5 years) visited the laboratory on two non-consecutive days. Following a randomized order, in one of the sessions participants completed a 3-min submaximal isometric contraction of the plantar flexors at 40% of maximal voluntary contraction (MVC), previously determined. Then, for testing, a cuff was inflated around the proximal region of the ipsilateral thigh to elicit metabolic entrapment via PECO (at 300 mmHg) beginning 5-10 s before exercise termination (cuff day). On the other session, testing was completed without previous metabolic accumulation and under free-flowing conditions (no-cuff day). The individual soleus H-reflex recruitment curve was obtained during a plantar-flexion tonic contraction (10% MVC), through tibial nerve stimulation, to test peripheral (maximal M-wave (Mmax)) and spinal (H-reflex maximal normalized amplitude (Hmax/Mmax), slope (H slope) and threshold (Hthresh)) excitability. Then, soleus electromyographic (EMG) amplitude and spinal/supraspinal (V-wave maximal normalized amplitude (V/Msup)) excitability were determined during additional MVCs. Pain rating (0–10-point scale) was self-reported and brachial arterial pressure was taken throughout experiments. Results: During PECO, MVC (-32.3 ± 17.5%; p = 0.03) and V/Msup (-42.0 ± 45.7%; p = 0.01) decreased when compared to the no-cuff day. Conversely, EMG amplitude (p = 0.17), Mmax (p = 0.19), H slope (p = 0.87), Hthresh (p = 0.57) and Hmax/Mmax (p = 0.62) remained unchanged. Pain rated + 6.0 points (p < 0.001) in the cuff day and systolic and diastolic arterial pressure were higher with than without ischemia (163.8 ± 93.1 vs. 134.8 ± 74.1 mmHg; p < 0.001). Conclusions: Continued input from type III/IV muscle afferents do not affect soleus H-reflex excitability, but leads to a decrease in maximal plantar-flexor torque output that is accompanied by a depressed descending neural drive. Ultimately, these data provide preliminary evidence that muscle afferents exert a modulatory effect at the supraspinal level to inhibit the activation of soleus efferent pathway.

Area 3 - Sport Performance and Support Technology

Nr: 64
Title:

Changes in Muscle Reoxygenation Post-Vascular Occlusion and Post-Exercise Representing Training Adaptation of Aerobic Capacity.
Authors:

Heru Syarli Lesmana, Patrick Rodrigues, Lydia Simpson, Kyohei Marume, Hendrik Mugele, Dean Perkins and Justin Lawley
Abstract: Purpose: Exercise training adaptation improves maximum aerobic capacity, and it could be explained by enhanced vascular function and skeletal muscle oxidation capacity. This study aimed to investigate whether improvement in hyperemic blood flow, muscle reoxygenation post-vascular occlusion, and post-exercise occur after a six-week training intervention to monitor training adaptation. Methods: Sixteen healthy participants, ten males and six females (age, 28.2±5.4 yr; stature, 1.77±7.6 m; weight, 76.7±13.6 kg; blood pressure, 130/80±17.4/9.4 mmHg) volunteered for 6-week of high-intensity interval training (HIIT). Before and after the intervention, participants underwent three different tests on different days: 1) a vascular occlusion test on the upper thigh to measure muscle reoxygenation (SmO2) with near-infrared spectroscopy and peak blood flow (PBF) via ultrasonography during reactive hyperemia (RH); 2) an incremental exercise test to assess parameters of maximum aerobic capacity including maximum oxygen consumption (V̇O2max), peak power output (PPO) and maximum heart rate (HRmax); and 3) a steady-state exercise (SSE) to assess post-exercise SmO2 recovery. The rate of relative SmO2 recovery back to baseline (Rbl) and reperfusion 10 seconds (Rep 10S) were calculated with the increment (I) divided by time (T) Results: The adaptations in maximal aerobic capacity were observed with increases of absV̇O2max (P<0.001; ES=0.65; 3.60±0.67 to 4.03±0.66 l/min), relV̇O2max (P<0.001; ES=0.66; 47.7±9.49 to 53.6±8.29 ml/kg/min), and PPO (P<0.001; ES=0.41; 300±47.1 to 320±48.9 watts) and reductions of HRmax (P=0.009; ES=0.35; 186±7.92 to 183±7.11 bpm). RH PBF increased post 6-week of training (P=0.04; ES=0.61; 1354±429 to 1487±281 ml/min). Improvements were also observed in SmO2 parameters post-vascular occlusion, including Rep 10S (P<0.001 ES=1.11; 0.74±0.28 to 1.17±0.45 %/s) and Rbl (P<0.001 ES=1.06; 1.27±0.45 to 1.94±0.78 %/s), linked to an increase in Ibl (P=0.03 ES=0.50; 10.4±4.57 to 12.9±5.11 %) and faster Tbl (P=0.03 ES=0.82; 8.45±2.03 to 6.96±1.61 s). SmO2 recovery post-SSE at the same workload for the pre-and post-test indicated slower recovery for Rep 10S (P<0.001 ES=0.33; 0.29±0.23 to 0.23±0.21 %/s) and Rbl (P=0.01 ES=0.27; 0.43±0.34 to 0.34±0.31 %/s) mediated by reduction of Ibl (P=0.01 ES=0.46; 5.87±4.37 to 4.06±3.41 %) and faster Tbl (P=0.01 ES=0.73; 14.5±3.38 to 12.1±3.09 s). Conclusion: After 6 weeks of HIIT intervention, increases in maximal aerobic capacity were seen alongside improvement in SmO2 post-vascular occlusion and efficiency in SmO2 post-exercise. Thus, the findings from this study indicated that changes in muscle reoxygenation potentially are an easy and practiced approach to monitoring training adaptation.
Nr: 63
Title:

Using Near-Infrared Spectroscopy to Describe Muscle Oxygen Saturation During Floor Performance in Collegiate Female Gymnasts
Authors:

Anna Campos, Julianna Dean and Julianna M. Dean
Abstract: INTRODUCTION: Muscle metabolism is often studied in athletes participating in traditional sports such as American football or basketball. However, there is little research in gymnastics and even less in female artistic gymnastics. Although gymnastics is usually classified as an anaerobic sport, the 90-sec floor routine demands contribution from the aerobic energy system. This contribution from the aerobic system is unstudied and can be a limiter of floor performance. PURPOSE: The purpose of this observational study is to describe muscle oxygen saturation (SmO2) levels in the vastus lateralis and observe if there is a relationship between SmO2 levels and scores on the floor exercise. We hypothesize that higher SmO2 values will correspond to higher average floor routine scores. METHODS: To be included, gymnasts must have participated in the 2023-2024 National Collegiate Athletic Association Division I (NCAA DI) season and were apparently healthy according to the American College of Sports Medicine guidelines. The head coach approached her team of 22 women; nine voluntarily consented. The study protocol included two days at the William & Mary Williamsburg Gymnastics training facility. On day one, consent forms were signed, we acquainted participants with the study, and we measured anthropometrics and demographics. On day two, the head coach decided the order of participation. Each gymnast performed their usual warm-up of about 15min. To the first gymnast, we attached a Moxy Monitor with Dynamic Tape to the vastus lateralis of her dominant leg. The Moxy Monitor uses near-infrared spectroscopy to measure real-time SmO2 at the capillary level. We synced the Moxy to PerfPro Studio software, which showed live SmO2 readings. The gymnast performed her floor routine. We then removed the Moxy from her and attached it to the next gymnast until all gymnasts were measured. ANALYSIS: We describe demographics and anthropometrics with counts, percentages, means, and standard deviations. We report average, minimum, and peak SmO2. We obtained each athlete’s 2023-2024 NCAA DI competition scores on the William & Mary public-facing website. We calculated an average of their floor scores throughout the season. RESULTS: Our sample characteristics were: 161.7 +/- 5.7cm tall, 63.0 +/- 7.3kg, 20.4 +/- 0.7yrs, 5 (55.6%) were right- and 4 (44.4%) were left-leg dominant, 1 (11.1%) was a first-year, 2 (22.2%) were second-years, 6 (66.7%) were third-years, and 0 (0.0%) were fourth-years. Since there were fewer than 10 participants, to maintain confidentiality, we report the average of those nine values for SmO2 results. Average SmO2 was 45.1 +/- 9.7%, the average minimum SmO2 was 30.7 +/- 11.8%, and average peak SmO2 was 63.9 +/- 6.3%. We found a positive trend where high values of average, minimum, and peak SmO2 trended toward higher floor scores throughout the competitive season. DISCUSSION: Our results are the first to describe muscle oxygen saturation levels in NCAA DI female gymnasts. The variation in SmO2 values in gymnasts may indicate various limiting factors of performance: cardiac, respiratory, or muscular oxidative capacity. Future research must identify what is predictive of top floor performance, e.g., consistently high SmO2, or the ability to sustain low levels of SmO2. These results can benefit athletic stakeholders to identify potential physiological limiters and tailor training protocols for maximum performance.
Nr: 66
Title:

Elite-Level Cooperation and Opposition Dynamics During Defensive Transitions: Using Computer Vision Data to Estimate the Pass and Dribbling Progression Conceded
Authors:

Rui Fernandes Freitas, Rui J. Lopes, Jani Sarajärvi and Anna Volossovitch
Abstract: This study examined the influence of emergent cooperation and opposition dynamics on ball progression permitted to opponents, during open-play defensive transitions in association football. Ninety-four matches from the top three ranked teams in the Portuguese First League season 2022/23 were included in our sample. Thirty-six variables related to the match-play landscape were analysed using Linear Mixed Models, to examine the predictors of ball progression through either passing, dribbling, or by a combination of both. Results indicated an inverse relationship between the first defender’s average distance to the ball and passing advances, suggesting that a greater distance deters opponents from progressing through passing. The variation in the first and second defenders’ distances to the ball showed mixed effects, with increased distances generally hindering progression, irrespective of its type. Voronoi cell analysis revealed that larger areas for the third defender were linked to greater spatial progresses, likely reflecting opponents’ success rather than defensive failures. Unexpectedly, team spatial dominance around the first defender was associated with less defensive success, possibly due to a lack of compactness and an episode’s selection bias. Increases in the angles between defenders and their own goal were positively correlated with all types of progression, emphasizing the importance of maintaining alignments to impede advances throughout the pitch. Overall, these findings reveal how certain defensive patterns, related to players’ relative distances, individual and collective areas of dominance and player-environment angles, could be adopted to better obstruct opponents’ ball progression, during open-play defensive transitions in football.
Nr: 68
Title:

Interpretable Low-Dimensional Modeling of Spatiotemporal Agent States for Decision Making in Football Tactics
Authors:

Kenjiro Ide, Taiga Someya, Kohei Kawaguchi and Keisuke Fujii
Abstract: Understanding football tactics is crucial for managers, coaches, and analysts. Previous studies have proposed mathematical models based on spatial evaluation and minimum distance to the ball using geometrical and kinematic equations, but these models are often computationally expensive. Recent reinforcement learning approaches utilize the positions and velocities of all players but suffer from low interpretability and require large datasets. Rule-based evaluation models address both limited data and alignment with expert knowledge, yet prior work did not consider the states of all players to define their actions. This study investigates whether low-dimensional, rule-based state models using spatiotemporal data can effectively capture and understand football tactics. Regarding passing actions, researchers have considered modeling and valuing of a pass in data-driven manners. Unlike these studies, our research adopts an approach that aims to understand decision-making in football tactics by performing interpretable low-dimensional modeling of spatiotemporal agent states in a rule-based manner. Our approach involves defining interpretable state variables for the ball-holder and multiple potential pass receivers. These state definitions are based on criteria that explore options such as passing and dribbling. For our analysis, we utilize StatsBomb event data and SkillCorner tracking data for the LaLiga 2023/2024 season. This data enables us to represent our defined state variables and tactical actions accurately. In the analysis, first, we identify important variables and train a XGBoost model to predict future goals. Validate the performance of the model using F1 scores with test data to see if the pass is successful or unsuccessful. The results showed that the distance between the player and the ball had a relatively large impact on the success of a pass. Our interpretable low-dimensional modeling made it easier to interpret the results by using intuitively understandable variables such as the distance between players and the ball, and space scores, in predicting pass success. Furthermore, the high interpretability allows for feedback from managers and analysts, potentially contributing to tactical improvements. Therefore, this interpretable low-dimensional model is considered to have practical value as a support tool for understanding and decision-making in tactical analysis on the field.
Nr: 70
Title:

Potential Predictors of Discrepancies Between Subjective and Objective Training Load in Elite Endurance Athletes
Authors:

Lina Spetz and Filip J. Larsen
Abstract: Introduction. Elite endurance sports require careful balancing of training load and recovery, where even minor disruptions may interfere with performance, development, health, and well-being. The monitoring regimen of load and recovery-related parameters is, therefore, a crucial component of daily life for elite endurance athletes. The increasing use of wearable technology provides continuous data streams on various physiological parameters, which can offer valuable insights for both athletes and coaches. However, easy access to objective data may lead to de-prioritization of subjective experience, potentially overlooking signs of imbalance, fatigue, or psychological stress. Even though exercise training is possible to objectively quantify, the perception of how burdensome sessions are experienced may differ, leading us to explore the ratio between subjectively rated and objectively calculated training load as a measure. Therefore, this study aimed to explore possible discrepancies between subjective and objective training load in elite endurance athletes, and to identify predictors that may influence the subjective-objective ratio. Method. We conducted an observational study with 10 male and 8 female elite endurance athletes, all competing for their respective national teams. During a full year of regular training and competitions, daily data collections were made to gain subjective and objective data with individual variance. Objective data were recorded using GPS watches and HR chest straps during training sessions and competitions, while subjective ratings (training load, muscle soreness, mental stress, sleep, food intake, mood, and energy level) were gathered via the Readiness Advisor (RA) app, with scores ranging from 0-10. First, we calculated the annual averages for RA and objective training load, followed by a Spearman correlation analysis of these. We then computed the ratio between subjective and objective load and created a multiple linear regression with the ratio as the dependent variable and the subjective RA ratings as predictors, both for individual athletes and at the group level. For each athlete, the RA parameter with the highest absolute coefficient was identified, and at the group level, the most frequently identified. Results. The initial correlation analysis revealed a non-significant negative association between perceived and objective load. Although not statistically significant, RA Mental stress was the strongest predictor of load ratio at the group level, with a coefficient of 0.06. This suggests that exercise training may be perceived as more burdensome than the objective training load indicates if mental stress is increased. At the individual level, RA Food was the most frequent predictor, with five athletes identified, while RA Mood followed with four athletes. However, which direction the prediction indicated varied for both RA Food and Mood between individuals. Conclusion. These findings underscore the importance of personalized monitoring, as group-level trends may not accurately capture individual athlete responses. A tailored approach is necessary to properly manage training load and recovery, reducing the risk of imbalance.
Nr: 71
Title:

Optimizing Performance with AI-Driven Adaptive Training Programs
Authors:

Olof Södergård, Lina Spetz and Filip J. Larsen
Abstract: Balancing training load with adequate recovery is a key challenge in optimizing athletic performance. This study investigated the potential of a machine learning-driven adaptive training program that used both objective data and subjective self-report metrics to optimize performance outcomes. Thirty-one recreational runners were enrolled in a 12-week longitudinal study, consisting of a 4-week baseline period followed by 8 weeks of intervention. Participants were randomly assigned to either an adaptive group (n=15), where an AI-generated training plan was dynamically updated based on accumulated training load and daily readiness ratings (captured via the Readiness Advisor app), or a control group (n=16) that followed a fixed training protocol throughout. Training load was continuously monitored using Garmin devices, and physiological performance markers, including submaximal and maximal running tests, running economy, threshold speed, and a 3000m time trial were collected at baseline and post-intervention. Results demonstrated that the adaptive group improved 3000m time trial performance by 22.3s while the static group improved by 14.2s (p=<0.01). Threshold speed increased marginally by 0.3 km/h in the adaptive group (p<0.05) vs. +0.2 km/h (n.s) in the static group. More notably, running economy in the adaptive group improved significantly (-3.5 ml O2/kg), in contrast to a deterioration in the control group (+4.5 ml O2/kg). The integration of real-time data and self-reported metrics via AI-driven algorithms enabled more responsive and effective modulation of training stimuli, underscoring the potential of adaptive training systems in enhancing key performance metrics. Conclusion: The study provides evidence that AI-based adaptive training, through continuous real-time adjustments informed by several data streams, enhances performance outcomes compared to static training models, particularly in parameters such as running economy. This approach represents a way to adjust training on a daily basis to better accommodate differences in athlete readiness and improve performance.