ECSS Paris 2023: OP-AP46
INTRODUCTION: Wearable sensors have emerged as a viable alternative to traditional laboratory equipment for assessing biomechanical parameters. This study investigated whether data from a 6-axis inertial measurement unit (IMU) mounted on a chest-strap, along with a deep learning model can replicate the ground reaction forces (GRF) measured by a force plate, potentially broadening access to performance and injury risk assessments in sports science. METHODS: 124 valid jumps from 12 participants with 2 years of resistance or endurance training experience were used for training. Synchronized recordings from a chest-mounted IMU and a force plate were obtained, with force plate data processed to extract the acceleration component of the GRF—normalized for body mass and with the gravitational component removed. Two deep learning architectures were employed: a vanilla auto-encoder and a recurrent auto-encoder incorporating gated recurrent unit (GRU) cells. Both models were trained for 500 epochs using 80% of the dataset, with the remaining data reserved for testing. Performance was evaluated using Mean Squared Error (MSE) as the loss metric. RESULTS: The base MSE loss on the test set was 0.3145. Vanilla auto-encoder achieved a training loss of 0.0021 and a test loss of 0.0717, while the GRU auto-encoder recorded a training loss of 0.0057 and a test loss of 0.0774. Both models effectively reconstructed the GRF-derived acceleration signal, with the vanilla architecture showing marginally better performance on unseen data, while also being faster for training and inference. CONCLUSION: These findings indicate that chest-mounted IMU data, when processed through an auto-encoder framework, can reliably approximate the acceleration component derived from GRF measurements. The comparable performance of both auto-encoder models suggests that even simple architectures can effectively translate the noisy data collected from a wearable sensor to GRF. This approach has the potential to facilitate on-field assessments where traditional force plates are impractical. Future research should focus on validating these preliminary results across varied athletic populations and movement patterns, as well as optimizing model parameters for real-time applications. Furthermore, additional pre-processing of the IMU signal may also lead to improved performance on this task, and requires further enquiry.
Read CV Tanuj WadhiECSS Paris 2023: OP-AP46
INTRODUCTION: The shoulder is the most injury-prone area in functional fitness, particularly during high-volume movements such as kipping pull-ups (KPUs) and butterfly pull-ups (BPUs) [1,2]. Overuse and overload have been associated with these injuries [1], making joint load monitoring a promising strategy for injury prevention. However, current methods rely on highly specialized equipment, and measuring interaction forces at the pull-up bar is challenging even in lab-based conditions, with limited practical applicability. Inspired by previous work in running [3], we propose using a minimal set of wearable sensors to estimate bar reaction forces during different pull-up techniques and we validate estimates against force plate measurements. This setup can contribute to field-based monitoring of shoulder joint loads across different pull-up techniques. METHODS: Nine experienced recreational CrossFit athletes performed one set of multiple upper body functional movements (Func), followed by sets of strict pull-ups (SPUs), KPUs, and BPUs. A pull-up rack was instrumented with two 6-DOF force transducers to measure the forces and moments at the handles. Participants wore two IMUs—one on the chest and one on the humerus. A two-step machine learning pipeline was developed. Step 1 used a multi-class classifier to distinguish between SPUs, KPUs, BPUs and other movements (Func). Step 2 applied a regression model for each technique, using raw IMU data to estimate anterior-posterior (Fap) and superior-inferior (Fsi) forces (medio-lateral forces were negligible as the movements were largely symmetric). Multiple machine learning models – including feature-based, kernel-based, and neural networks – were tested using single-sensor (chest only) and dual-sensor configurations. Validation used a leave-one-subject-out approach to ensure generalizability. RESULTS: For movement classification, the best-performing model, Rocket classifier, single sensor, achieved 98±2% accuracy and F1 scores. In estimating forces normalized by body mass, the best model, LSTM, dual-sensor, yielded RMSE [Fsi, Fap] in N/kg: KPU: [0.30, 0.15], BPU: [0.27, 0.11], and SPU: [0.13, 0.05]. Normalized by the full range of values, these are [Fsi, Fap]: KPU: [21%, 17%], BPU: [15% 11%] and SPU: [16%, 22%]. Linear coefficients of determination were [Fsi, Fap]: KPU: [0.70, 0.75], BPU: [0.83, 0.90], and SPU [0.59, 0.33]. Differences between measured and predicted peak forces varied largely between trials and participants. CONCLUSION: We present a data-driven method to estimate external forces during pull-ups using 1-2 wearable sensors with strong accuracy. This setup holds promise for field-based biomechanical assessments which can be used for shoulder load monitoring, preventing overload and overuse injuries. Future work will evaluate how accurate these estimated forces can be used to compute joint moments via musculoskeletal modeling and inverse dynamics. 1. Dominski et al. (2022) 2. Summitt et al. (2016) 3. Scheltinga et al.
Read CV Mariah SabioniECSS Paris 2023: OP-AP46
INTRODUCTION: Ventilatory thresholds (VTs) mark aerobic and anaerobic metabolic transitions in exercise physiology. Traditionally, detecting these thresholds during cardiopulmonary exercise testing (CPET) involves manual assessment of respiratory gas exchange data based on visual inspection, which is time-consuming and prone to subjectivity. To address these limitations, we developed an algorithm to automate the detection of VTs and conducted an initial analysis of its performance based on the Power at which VTs were reached. METHODS: An algorithm was developed in Python (Version 3.13.1) to detect VT1 and VT2 from continuous CPET data based on established criteria. VT1 was detected using four parameters: (1) the ventilatory equivalent for oxygen starts rising, (2) end-tidal oxygen pressure starts rising, (3) the V-slope method, and (4) a rise in excess carbon dioxide. VT2 was detected using three criteria: (1) the ventilatory equivalent for carbon dioxide starts rising, (2) end-tidal carbon dioxide pressure starts decreasing, and (3) the V-slope method. The algorithm was applied to respiratory gas exchange data from 210 participants under 50 years old in the COmPLETE-Health study [1]. VTs detected by the algorithm were compared to visually determined values, with performance assessed using Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) for Power at VT1 and VT2. RESULTS: The algorithm successfully detected VT1 in 92.9% of participants and VT2 in 93.3%, respectively. Among participants with valid threshold detections, the MAPE for VT1 was 12.6%, with an RMSE of 14 Watt, while VT2 had a MAPE of 5.7% and an RMSE of 11 Watt. CONCLUSION: The observed error rates indicate a reasonable agreement with visual determination for both VT1 and VT2, respectively. Considering that an automated detection of VTs is i) more objective, ii) better reproducible, and iii) more time and cost efficient than manual visual VT determination, this provides high potential. 1. Wagner et al. 2021. Novel CPET reference values in healthy adults: associations with physical activity. Medicine & Science in Sports & Exercise, 53(1), 26-37.
Read CV Laura StützECSS Paris 2023: OP-AP46