A scoping review auditing 258 full-text studies has found that predictive modelling and artificial intelligence in team-sport athlete monitoring suffer from substantial validation and reporting deficits that limit their trustworthy prospective use. Although the application of machine-learning and conventional statistical forecasting methods has surged across athletic environments, the researchers identified that empirical validation protocols routinely fail to replicate real-world deployment conditions. The investigation, led by Šimon Salaj, Michal Hrubý, and Jan Cacek from the Faculty of Sports Studies at Masaryk University in Brno, Czechia, mapped literature published between 2010 and 2026 to evaluate whether published sports-monitoring models meet foundational standards for trustworthy implementation.
To evaluate the scientific literature, the researchers conducted systematic database searches across OpenAlex, PubMed, Scopus, and SPORTDiscus covering records from 2010 through 2026. Following deduplication, initial screening, and an updated search rerun in OpenAlex, the team identified 258 studies meeting inclusion criteria for full-text auditing. The methodology noted specific procedural constraints: systematic duplicate screening was not performed, and initial eligibility determinations along with original audit coding were carried out by a single reviewer rather than an independent screening team.
The included investigations were evaluated using a specialized ten-domain audit framework designed to assess trustworthy artificial intelligence practices. This instrument examined prediction context, temporal validity, repeated-observation independence, external validation, leakage control, calibration, class-imbalance handling, interpretability, uncertainty reporting, and practical deployment or actionability. In addition to manual auditing, the authors applied bibliometric mapping and natural language processing analyses to examine broader publication characteristics, data modalities, modelled outcomes, and model families across the discipline. To evaluate assessment reliability, a second reviewer independently evaluated 45 studies drawn from the initial 223-study subset across all ten audit dimensions.
Analysis of publication volume showed that literature in multimodal load-response modelling expanded rapidly after 2021, with athletic performance modelling and injury-risk prediction representing the dominant applied objectives. Monitored athletes were assessed through diverse diagnostic streams, most frequently incorporating global positioning and global navigation satellite systems (GPS/GNSS), session ratings of perceived exertion (session-RPE) or internal load metrics, wellness and sleep questionnaires, heart rate and heart-rate variability (HR/HRV), neuromuscular performance tests, inertial measurement units (IMU/accelerometry), and video-derived kinematic features.
Regarding computational architectures, the audit showed that random forests, logistic and linear regression formulations, and gradient boosting represented the most frequently deployed model families.
The authors found that independent external validation was strongly documented in just one single study (0.4 percent) and partially documented in one additional investigation (0.4 percent). Other fundamental machine-learning safeguards were likewise uncommon across the corpus, with studies rarely demonstrating adequate control for data leakage, rigorous probability calibration, adjustment for repeated-observation non-independence among individual players, or comprehensive model-level uncertainty reporting.
Reliability testing within the independent two-reviewer audit subset revealed a pre-correction exact agreement of 65.3 percent across the evaluated studies, with agreement metrics displaying substantial variation across the individual audit domains. The authors concluded that predictive model development in team-sport athlete monitoring has progressed substantially faster than trustworthy validation practice. To address these systemic gaps, they recommended that prospective research clearly specify the operational decision context, maintain underlying temporal and player-level data structures, enforce strict leakage controls, systematically document calibration and prediction uncertainty, validate systems under authentic field conditions, and deliver actionable decision-support tools for practitioners.
Subject of Research: Trustworthy artificial intelligence validation practices in multimodal team-sport athlete monitoring
Article Title: Trustworthy AI for multimodal load-response modelling in team sports: a scoping review and full-text validation audit
Article References: Salaj, Š., Hrubý, M., & Cacek, J. (2026). Trustworthy AI for multimodal load-response modelling in team sports: a scoping review and full-text validation audit. Frontiers in Sports and Active Living, 8, 2001654. https://doi.org/10.3389/fspor.2026.2001654
Image Credits: AI Generated
DOI: 10.3389/fspor.2026.2001654
Keywords: athlete monitoring, team sports, scoping review, trustworthy AI, predictive modelling, load-response modelling, validation audit, temporal validity, external validation, data leakage control, model calibration, repeated-observation independence, uncertainty reporting, injury-risk prediction, performance modelling, random forests, gradient boosting, session-RPE
News Source: Ophelia Keating. (November 11, 2026). Audit Finds Rigorous Validation Is Rare in Team Sport Predictive AI. Scienmag.



