Scientific Programme
Applied Sports Sciences
IS-AP05 - AI and mathematical models applied to sport from research to real-world applications
Date: 09.07.2026, Time: 08:30 - 09:45, Session Room: Auditorium A (STCC)
Description
The role of the sports scientist encompasses the collection, analysis, interpretation, presentation and protection of data. Modeling provides the most in-depth understanding of the phenomena, although biomechanical and physiological modeling applications in sports science remain underexplored. Recently, the substantial growth of AI-based models in sports science has opened unprecedented opportunities, making cutting-edge analytics available to researchers, analysts, and practitioners. Yet the human contribution gives meaning and value to analyses and models. The use of AI-based analytics through systematic frameworks provide interpretable metrics that inform decision-making for athletes, coaches, medical staff, and front offices, thereby yielding benefits in prediction of sports match outcomes, talent scouting, tactical analysis, injury risk reduction, and optimisation of sports performance. In team sports, advanced methods include counterfactual estimating and comparing observed behavior to computed optima. There have been efforts made to make advanced tools available. Despite the proceedings, concerns persist regarding accessibility, transparency and design for both academic studies and real-world applications. This symposium is designed to serve as a nexus between academic research and real-world sports applications with the aim of establishing a complete framework for creating and applying AI-based analytical models in sports science.
Chair(s)
Andrea Zignoli
University of Trento, Department of Industrial Engineering
Italy
Speaker A
Andrea Zignoli
University of Trento, Department of Industrial Engineering
Italy
Read CVECSS Rimini 2025: IS-AP05
From First Principles to Artificial Intelligence: Rethinking Modelling in Sports Science
Modelling is the most fundamental act of science: the process of constructing simplified yet meaningful representations of reality. In sports science, every measurement, equation, or simulation is a model of something that happens in the body, in performance, or in decision-making. Yet as the field increasingly turns to artificial intelligence (AI) and data-driven approaches, the foundations of modelling are often overlooked or treated as purely technical. In this talk, I will revisit the basics of modelling: why we model, what a model is, and how models “think”. Through three examples and lines of research, I will try to reframe their relevance for practitioners, researchers, and stakeholders in the evolving landscape of sports science. First, I present how first-principles models can be used to examine cornering and pacing strategies in road cycling. By deriving relationships between power, speed, and physical constraints, we can simulate how athletes distribute effort and optimize trajectories under physical laws. These models capture the logic of performance and motor control, revealing how athletes implicitly “think” through physics. Second, I present Oxynet and how we can integrate AI technologies to detect patterns in ventilatory variables during cardiopulmonary exercise testing. By automatically identifying transitions between exercise intensity domains, it becomes a digital lens through which we can observe physiological regulation. Third, I will explore semantic analysis to compare AI-driven and human endurance sport coaches, introducing one of the most disruptive technologies of our time: the large language models. By mapping linguistic and conceptual pathways, we trace the trajectories of AI and human coaches' reasoning and reveal what aspects of coaching (e.g. empathy, contextual judgment, and intuition), remain uniquely human. These cases illustrate different faces of the same modelling world: how humans think with equations and how machines “think” with networks. For machines, modelling remains confined to computation and data retrieval, but for sports scientists, it is a cognitive and interpretative process. It is up to the sports science community to regulate the use of these tools, so the value of human contribution that gives meaning to data is preserved and valorised, to ensure that professionals are empowered to leverage AI ethically and intelligently. Only by doing so can we preserve the essence of modelling as a bridge between understanding and action, not as a replacement for human thinking, but as its extension.
Speaker B
Lorena Martin
USC Marshall School of Business, Department of Data Sciences and Operations
United States
Read CVECSS Rimini 2025: IS-AP05
Progressing Protocol to Platform: Engineering the Future of Sport through Artificial Intelligence
Artificial Intelligence (AI) has opened unprecedented opportunities for sport scientists to interpret athlete data with greater accuracy and scale. Yet, its effectiveness depends on the integrity of the underlying data processes. This work builds upon The Sports Science Data Protocol (Martin, 2019), a systematic framework for identifying key performance indicators (KPIs), validating technologies, ensuring reliability, securing consent, and presenting findings — and demonstrates its evolution into an AI-enabled ecosystem now operational in professional basketball, baseball, tennis, and soccer. The resulting platform, SportsVision.ai, integrates computer vision and large-language-model (LLM) technologies to analyze video and sensor data, addressing three pervasive challenges in elite sport: injury prevention, in-game strategy optimization, and scouting accuracy. By automating data ingestion, validation, and feature extraction, SportsVision.ai transforms raw performance footage into interpretable metrics that inform decision-making for athletes, coaches, medical staff, and front offices. Applications include predictive modeling of biomechanical micro-deviations to anticipate injuries, reinforcement-learning systems for tactical optimization, and objective video-based scouting that reduces bias and enhances player valuation. Across the NBA and comparable leagues, these implementations have demonstrated potential to mitigate annual losses of $40–125 million per franchise due to injuries, inefficiencies, and poor evaluations. By operationalizing a validated research protocol within an AI framework, this work establishes a replicable, ethical, and explainable model for integrating artificial intelligence into sports performance science. The result is a bridge from research to real-world application—where data integrity, athlete trust, and computational intelligence coalesce to redefine how performance is measured, protected, and optimized.
Speaker C
Keisuke Fujii
Kyoto University, Graduate School of Human and Environmental Studies
Japan
Read CVECSS Rimini 2025: IS-AP05
Development and Democratization of Artificial Intelligence in Team Sports
Rapid advances in sensing, image processing, and machine learning have raised expectations for AI-driven analysis in sports. At the same time, practical barriers remain: rights and access to data, large variations in camera setups, and the intrinsic complexity of multi-agent movement and decision making. These factors make reliable data collection and principled modeling of team behavior challenging. This talk presents a set of methods and systems aimed at addressing those challenges for team sports. First, I describe image-based pipelines for automatic data acquisition that combine multi-object tracking, identity re-identification, pose estimation, and geometric calibration to produce game-level tracking and event data from broadcast and fixed-camera video. Second, I discuss predictive models for counterfactual evaluation, which estimate what would have happened under alternative actions and thereby offer an interpretable way to assess individual and collective contributions. Third, I report on work using multi-agent reinforcement learning to produce valuation functions and action recommendations that span whole-game phases and all players, allowing comparison of observed behavior to computed optima. Beyond methods, I outline efforts to make these tools widely available. I summarize open releases of diverse datasets, the OpenSTARLab open-source analysis platform, and community competitions that aim to lower the entry barrier for researchers and practitioners. Finally, I indicate promising directions for applied research and collaboration.