Raket Sporlarında Giyilebilir Sensörler: Yüksek Bildirilen Doğruluk, Ancak Gerçek Dünya Performansı Belirsiz

Özet (abstract)

Wearable sensors are increasingly used for stroke and action classification in racket sports, yet reported performance may reflect not only model capability but also sensor placement, validation design, participant-level dependence, and potential data leakage. This systematic review and meta-analysis synthesized reported testing-instance-level performance of wearable sensor-based stroke and action classification systems in racket sports and explored the associations of biomechanical and methodological factors with reported accuracy. Web of Science, Scopus, PubMed, and IEEE Xplore were searched from database inception to May 1, 2026. Eligible studies used wearable or portable sensors for stroke and action classification in racket sports and reported extractable classification performance. Reported classification accuracy was transformed into chance-corrected accuracy and then logit-transformed for random-effects meta-analysis. Exploratory subgroup analyses and random-effects meta-regression were conducted according to sensor placement, algorithm category, validation strategy, and leakage risk. An additional sensitivity analysis was restricted to studies in which the extracted performance estimate was derived from explicit participant-independent separation between the training and testing datasets. Twenty-one studies, contributing one eligible effect estimate each, were included in the quantitative synthesis. The pooled logit-transformed chance-corrected accuracy was 3.052 (95% CI: 2.677–3.426), corresponding to a back-transformed chance-corrected accuracy of approximately 95.5% (95% CI: 93.6%–96.9%). Between-study heterogeneity was substantial ( \(\:{I}^{2}\) = 94.87%), and the 95% prediction interval ranged from 79.3% to 99.2% on the back-transformed chance-corrected accuracy scale, indicating considerable uncertainty in the expected performance of a future comparable study. The pooled estimate should therefore be interpreted as a summary of reported testing-instance-level performance under heterogeneous validation procedures, rather than as an estimate of accuracy in previously unseen athletes or real-world deployment. Five studies met the participant-independent criterion. Their pooled chance-corrected accuracy was 90.0% (95% CI: 81.9%–94.7%), although between-study heterogeneity remained very high ( \(\:{I}^{2}\) = 97.10%). The participant-independent pooled estimate was lower than the overall pooled estimate, indicating that performance may be reduced when models are evaluated in previously unseen participants. Validation strategy showed an exploratory between-subgroup difference, with participant-independent evaluation yielding the lowest pooled estimate. Algorithm category and leakage risk did not show statistically significant subgroup effects. Sensor placement showed an exploratory subgroup signal favoring racket-only configurations, but its omnibus meta-regression test was not statistically significant. All moderator findings should be interpreted cautiously because subgroup sizes were small and substantial residual heterogeneity remained. Wearable sensor-based systems show high reported testing-instance-level performance for racket-sport stroke and action classification. Performance remained relatively high in studies using participant-independent evaluation but was lower than the overall pooled estimate. The limited number of participant-independent studies and very high heterogeneity indicate that generalizability to previously unseen athletes remains uncertain. Potential priorities for future research include participant-independent, cross-session, and external validation, larger and more diverse samples, standardized reporting, and transparent training–testing separation.

Yazarların özeti; kaynağından alınmıştır. BMC Sports Science Medicine and Rehabilitation, 2026 · DOI ↗