Creative Commons source reading · Prediction and forecasting
Predictive frameworks for youth basketball: a leakage-aware reassessment of athlete development and injury-risk models
Chen Yao
Originally published in Front Physiol · 2026-10-22
From the article
Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.
Abstract
IntroductionArtificial intelligence is increasingly applied in sports analytics; however, exceptionally high predictive performance may result from target construction, data leakage, or deterministic label structures. This study reassessed predictive models for youth basketball using a leakage-aware framework.MethodsA publicly available youth-basketball dataset containing 1,000 records and 25 variables was analyzed. Development-score construction and injury-risk label structure were audited. Construction-related predictors were subsequently excluded, and model performance was evaluated using repeated cross-validation and a locked 200-athlete holdout. Subgroup analyses by gender and age and an external-validation feasibility assessment were also conducted.ResultsDevelopment_Score was reconstructed almost perfectly from its documented…
Results
The reviewer-driven reanalysis changed the interpretation of the original results. The near-perfect Development_Score and all-predictor injury results were reproducible, but both were explained by target construction or deterministic label structure. The principal evidence therefore comes from leakage-aware sensitivity analyses, locked-holdout performance, external-validation feasibility, and subgroup diagnostics.
Theoretical and practical implications
The principal implication is methodological. Sports-analytics studies can report excellent discrimination or fit while still offering little independent predictive information if the target is calculated from inputs or labels follow deterministic rules. In the present dataset, overall accuracy of 0.750 after proxy exclusion might appear acceptable in isolation, yet balanced accuracy of 0.329, macro-F1 of 0.286, and zero High-risk sensitivity reveal severe failure on clinically or operationally important minority classes. This illustrates why class-sensitive metrics, locked evaluation, and target-definition audits are necessary alongside headline accuracy or AUC. Practically, the current model should not be used to automate injury-prevention, recovery, or training-intensity decisions. Future athlete-management applications require…
Editorial reading guide
- What data created and tested the model, and was evaluation independent?
- What is the cost of a false positive, false negative or out-of-distribution decision?
