Key result
Deep learning with multimodal wearable sensors beats traditional models for injury prediction, achieving ~91% accuracy.
Why the study?
The rapid development of wearable devices and AI in sports science required summarization to identify methodological gaps and technical limitations hindering clinical adoption for injury prevention and performance optimization.
Does artificial intelligence and wearable technology improve injury prediction and load management in athletes?
Systematic Review (n=87)
Does artificial intelligence and wearable technology improve injury prediction and load management in athletes?
Deep learning models using multimodal wearable sensor data show high accuracy for athlete injury prediction, though real-world implementation is currently limited by data heterogeneity and a lack of longitudinal validation.
Answers are grounded in this paper and cite their sources.
Captured external expert commentary on this paper, strongest first. Original sources are linked where available.
“This is a broad evidence map, but performance inside retrospective datasets is not equivalent to prospective clinical utility. Many models risk overfitting, and very few studies demonstrate that acting on predictions reduces injury or improves performance. Wearables may help organize workload, sleep, cardiovascular, movement, and subjective data. They should not be treated as diagnostic instruments or autonomous readiness systems.”
Deep learning models outperform traditional ones for wearable injury prediction; extends performance evidence but leaves open real-time clinical adoption due to heterogeneity and interpretability barriers.
Background The emergence of sports science is marked by a change in the integration of the artificial intelligence segment to examine the high-rate burst of development of the wearable devices that is needed to decipher functional, physiological, and biomechanical areas of interest. The given review of literature performs the analysis of the status of artificial intelligence to summarize this information to facilitate proactive injury prevention and performance optimization. Purpose In this study, an attempt will be made to list the systematic, categorized approaches of existing predictive modelling solutions using AI, to trace the key methodological gaps and technical limitations to embark on the path of clinical adoption of AI-grounded predictive models. Methods This review is based on strict adherence to the PRISMA principles and includes a thorough search of databases such as PubMed, Web of Science, and IEEE Xplore with the help of stringent requirements of including high-quality research. Findings Deep learning (CNN, ResNet) models are superior to traditional models in the range of 87–91% and ROC-AUC up to 0.94 (analysis of 87 studies). Better predictive power was achieved with multimodal sensor configurations than with single sensor configurations. But the high data heterogeneity, class imbalance, and poor interpretability are major hurdles to real-time clinical implementation. Conclusion The critical issues have been identified in this review, including inter-subject variability and class imbalance, and the need to standardize data acquisition protocols as a way of reducing the effect of feature noise on the generalizability of the model. Finally, this summary shows that longitudinal validation studies are desperately needed that do not focus on individual data sets to confirm whether AI-informed interventions are effective in high-performance sports settings.
No takes yet. Share an insight, caveat, or question.
Erzhuo Huang (2026) conducted a systematic review in Athlete load management and injury prediction (n=87). Deep learning models and multimodal wearable sensors vs. Traditional machine learning models and single sensor configurations was evaluated on Predictive accuracy and ROC-AUC for injury prediction. Deep learning models using multimodal wearable sensor data achieved superior injury prediction performance with 87-91% accuracy and ROC-AUC up to 0.94 compared to traditional machine learning models.
Synapse has enriched 10 related papers. Consider them for comparative context: