Methodological trends in machine learning for sport

Abstract

Since 2020, the field of machine learning research in sports has expanded. This includes predictive modelling for sports medicine and athlete monitoring using wearables and screening data, sequential modelling for trajectory prediction, and computer vision for broadcast understanding (event spotting, multi-object tracking, pose estimation). Peer-reviewed journal and major conference proceeding evidence published between 2020-01-01 and 2026-02-11 is integrated in this review (search date: 2026-02-11). After outlining repeatable search terms and inclusion criteria, we offer a thematic synthesis arranged by task family (recognition/segmentation, tracking and pose, tactical decision support, biomechanics/engineering prediction, and health/injury risk) and data modality (wearables, clinical/screening data, video, and tracking). Algorithm/sport frequencies throughout the extracted sample are reported in a quantitative synthesis that summarizes representative studies. The findings show that: (i) open benchmarks like SoccerNet-v2 (~300k annotations over 500 broadcast soccer videos) and FineGym (708 hours; 530 fine-grained elements) have fuelled the dominance of deep learning in video and pose pipelines; (ii) tree-based boosting and SVMs continue to be common for structured sports medicine data; and (iii) interpretability tooling (SHAP, permutation importance) is being used more and more to convert models into insights that coaches and clinicians can use. Restrictions include partial PRISMA count reconstruction and limited access to certain paywalled indexing services. We wrap up by discussing unresolved issues with deployment, fairness, domain shift, generalization, and label quality.