Scholarlink

<p>Traditional talent identification in sprinting relies heavily on chronological performance metrics, which are often confounded by biological maturation and the Relative Age Effect (RAE). While machine learning (ML) offers advanced pattern recognition capabilities, frameworks integrating technical and psychological data require rigorous verification to ensure practical validity. This study developed and internally validated a multimodal ML framework to predict future national-ranking attainment in adolescent sprinters, comparing its efficacy against traditional univariate models. A prospective cohort study was conducted with 500 adolescent sprinters (mean age 15.2 ± 0.8 years) from Ethiopia, Kenya, and South Africa. Baseline assessment included 2D kinematic analysis (smartphone video, 240 fps), neuromuscular capacity tests (Countermovement Jump [CMJ], Broad Jump), and psychological profiling (Grit, Growth Mindset). Socioeconomic status (SES) and Relative Age were included as covariates. The primary outcome was the binary attainment of a Top-10 National U20 ranking at Year 4. An XGBoost classifier was trained using 5-fold cross-validation. Retention was 97.2% (<i>n</i> = 486). The multimodal model achieved an Area under the Curve (AUC) of 0.82 (95% CI: 0.74–0.88), outperforming a baseline model using only Baseline 100&#xa0;m Time (AUC 0.61; <i>p</i> &lt; .001). Feature importance analysis identified Ground Contact Time (GCT) variability and Grit as the strongest predictors. While the model successfully identified successful athletes across different birth quarters, the low SHAP importance of Birth Quarter suggests the model relies heavily on downstream technical phenotypes rather than relative age directly. However, due to its limited precision (48.0%) and the small absolute number of positive outcomes in the hold-out test set (<i>n</i> = 15), the model cannot function as an absolute selection gatekeeper. Furthermore, the lack of direct biological maturation assessment remains a key limitation. This multidimensional approach demonstrates exploratory utility for early regional talent screening, though findings regarding RAE mitigation must be interpreted cautiously as an exploratory screening tool rather than a fully validated talent-identification framework.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A multimodal machine learning framework outperforms traditional performance metrics for predicting elite national ranking attainment in adolescent sprinters

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月7日

查看全文
  • Arefayne Mesfen Dessye

摘要

Traditional talent identification in sprinting relies heavily on chronological performance metrics, which are often confounded by biological maturation and the Relative Age Effect (RAE). While machine learning (ML) offers advanced pattern recognition capabilities, frameworks integrating technical and psychological data require rigorous verification to ensure practical validity. This study developed and internally validated a multimodal ML framework to predict future national-ranking attainment in adolescent sprinters, comparing its efficacy against traditional univariate models. A prospective cohort study was conducted with 500 adolescent sprinters (mean age 15.2 ± 0.8 years) from Ethiopia, Kenya, and South Africa. Baseline assessment included 2D kinematic analysis (smartphone video, 240 fps), neuromuscular capacity tests (Countermovement Jump [CMJ], Broad Jump), and psychological profiling (Grit, Growth Mindset). Socioeconomic status (SES) and Relative Age were included as covariates. The primary outcome was the binary attainment of a Top-10 National U20 ranking at Year 4. An XGBoost classifier was trained using 5-fold cross-validation. Retention was 97.2% (n = 486). The multimodal model achieved an Area under the Curve (AUC) of 0.82 (95% CI: 0.74–0.88), outperforming a baseline model using only Baseline 100 m Time (AUC 0.61; p < .001). Feature importance analysis identified Ground Contact Time (GCT) variability and Grit as the strongest predictors. While the model successfully identified successful athletes across different birth quarters, the low SHAP importance of Birth Quarter suggests the model relies heavily on downstream technical phenotypes rather than relative age directly. However, due to its limited precision (48.0%) and the small absolute number of positive outcomes in the hold-out test set (n = 15), the model cannot function as an absolute selection gatekeeper. Furthermore, the lack of direct biological maturation assessment remains a key limitation. This multidimensional approach demonstrates exploratory utility for early regional talent screening, though findings regarding RAE mitigation must be interpreted cautiously as an exploratory screening tool rather than a fully validated talent-identification framework.