Artificial Intelligence in Sports Science: Evidence Maturity, Athlete Digital Twins, Multimodal Analytics, and a Translation Agenda for 2026–2035
TL;DR
The strongest evidence came from narrowly defined tasks with explicit reference standards, including motor-skill assessment, biomechanical estimation, markerless motion capture, and athlete tracking, while evidence for injury prediction, wearable systems, generative AI, and athlete digital twins was less mature.
Abstract
rtificial intelligence is increasingly used in sports science, sports medicine, athlete monitoring, physical education, biomechanics, coaching, and performance analysis. However, many studies report technical success without demonstrating whether an approach is transportable beyond the development setting, useful in professional practice, or adequately protective of athlete interests. This systematic scoping review examined peer-reviewed research published from January 2016 to May 1, 2026 on machine learning, deep learning, computer vision, wearable and multimodal analytics, large language models, generative artificial intelligence, and athlete digital twins. Six bibliographic databases were searched, with Google Scholar and citation tracking used as supplementary discovery methods. The search identified 1,132 records; 901 records underwent title and abstract screening, 287 reports were assessed for eligibility, and 32 publications were included. Primary empirical and technical evidence was examined separately from review, conceptual, perspective, and governance literature. Ten fully verified primary empirical studies enabled direct study-level comparison. The strongest evidence came from narrowly defined tasks with explicit reference standards, including motor-skill assessment, biomechanical estimation, markerless motion capture, and athlete tracking. Evidence for injury prediction, wearable systems, generative AI, and athlete digital twins was less mature. Recurring limitations included small or homogeneous samples, single-dataset validation, limited calibration, incomplete reporting of model behaviour, restricted reproducibility, and a lack of prospective or independent evaluation. None of the included publications demonstrated sustained decision impact or attributable improvement in athlete, educational, sports-medicine, or organisational outcomes. Progress will require temporal and external validation, calibrated uncertainty, transparent reporting, representative datasets, interoperable multimodal systems, human oversight, and safeguards for athlete autonomy, privacy, fairness, and contestability.
