Framework for modelling and prediction of athletic game performance and injury risk

Abstract

Integrating Internet technologies and advanced software/hardware systems, including wearable devices, cameras, and analytics tools, has transformed data acquisition and analysis in sports, particularly basketball coaching. This revolution enables precise team composition, strategy formulation, skill assessment, athlete profiling, performance prediction, tailored training plans, and real-time adjustments during games. Collegiate basketball presents unique challenges, such as varied training loads, disturbed sleep and recovery cycles, hectic travel schedules, academic demands, and psychological stressors, all of which influence athlete performance, readiness, and injury risk. This research proposes a multi-modal framework employing machine learning (ML) and computer vision (CV) to model and predict game performance and injury risk, integrating var ious data modalities—such as sleep and recovery data from WHOOP straps, training load data via session ratings of perceived exertion, subjective stress and recovery metrics from the short re covery short stress (SRSS) questionnaire, vertical jump data using dual force plates, practice and game metrics from the polar team pro system, high-resolution vertical jump videos, travel data from game schedules, game performance metrics including game score and player efficiency rating (PER), injury data, and opponent data from National Collegiate Athletics Association (NCAA) Division I women’s basketball teams. The framework utilizes explainable AI (xAI) techniques to interpret model outcomes, providing feedback to coaches through an interactive dashboard and improving decision-making. This study was conducted in collaboration with Sacred Heart University’s women’s basket ball team, involving a total of 27 different athletes [age = 21.00 [±3.00] years, height = 174.21 [±19.27] cm, body mass = 73.98 [±11.52] kg] for 4 seasons (2020-24). Beginning in Season 1 (2020-21), we started collecting data using WHOOP straps for sleep and recovery metrics, weekly training assessments using the perceived exertion session rating (sRPE), and biweekly cognitive state evaluations using the SRSS questionnaire. We systematically record game performance, injury occurrences, and in-game statistics. We developed predictive models for game performance and injury risk using ensemble boosting with decision tree regressors and classifiers, integrating model-agnostic techniques for interpreting outcomes. We expanded our methodology in Season 2 (2021-22) by incorporating jump data from dual force plates and polar team pro systems. We analysed and predicted game performance at the athlete, team, and conference levels. In addition, we develop a CV-based framework for biome chanical evaluation of jump landings, enhancing injury risk prediction, and guiding coaching strategies. In Season 3, we focussed on athlete-specific variations in data to account for intra-athlete variability in modelling and predicting athletic readiness in terms of reactive strength index modified (RSImod). We introduced the concepts of athletic signature and smallest worthwhile change (SWC) for addressing the intra-athlete variability and improve accuracy. In Season 4 (2023-24), our research delved deeper into the analysis of biomechanical features extracted from countermovement jumps (CMJs) using advanced CV techniques. In addition, we conducted a detailed analysis of opponents to improve game strategy and player performance predictions. We developed deep learning models for automated injury risk detection based on biomechanical er rors observed in jump performance videos. This integrated approach aimed to optimise training strategies, predict game performance, and minimise injury risks through innovative methodolo gies and data-driven insights throughout all seasons. This analysis offers significant benefits to athletes and coaches alike. Athletes benefit from personalised insights into their sleep, recovery, and biomechanical performance, enabling them to optimise training and improve overall performance. Coaches gain valuable tools for strategic decision-making, including predictive models for game performance and injury risk, thus im proving team outcomes and player management. In addition, expanding these frameworks to other sports promises to revolutionise training methodologies and injury prevention strategies across diverse athletic disciplines.