Analyzing Workload-Monitoring Consistency and Same-Week Injury-Classification Performance in Female Adolescent Swimmers: An Exploratory Session-RPE and Neural Network Study

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Objective: This exploratory study aimed to: (1) verify whether session-RPE (sRPE) can serve as a reliable workload-monitoring indicator in female adolescent swimmers aged 13–17 years; (2) compare the same-week injury-classification performance of sRPE-, exponentially weighted moving average (EWMA)-, and robust exponential decreasing index (REDI)-derived indicators using a multilayer perceptron (MLP) model; and (3) identify candidate indicator combinations for future prospective validation in youth swimming. Design: A 12-week prospective observational tracking study was conducted with 20 female adolescent 200 m individual medley swimmers (aged 13–17 years) from a single provincial youth swimming team (Zhejiang). Twenty-five injury events were recorded across 240 athlete-weeks. Because repeated athlete-weeks were analyzed as independent samples and injury status was classified within the same week as the load indicators, the classification findings are preliminary internal-validation results that may overestimate model performance; they are not externally validated prospective predictions. Methods: Pearson and Spearman correlation analyses, intraclass correlation coefficient (ICC), and Bland-Altman plots were used to evaluate associations and agreement between sRPE-derived indicators and traditional training-load indicators (swimming distance and training duration). Twelve sRPE-derived indicators (four each from sRPE, EWMA, and REDI) were used as input features for MLP same-week binary classification models. REDI was included because its natural-logarithmic weighting emphasizes long-term cumulative load, complementing the recency-weighted EWMA. Model performance was evaluated using area under the ROC curve (AUC), with Bootstrap resampling (1000 iterations) to characterize internal stability. Results: Swimming distance, which shares no computational component with sRPE, was very strongly correlated with sRPE (r = 0.965, p < 0.001). Training duration was also strongly correlated with sRPE (r = 0.972, p < 0.001), but this duration correlation is partly part-whole by construction because duration enters the sRPE calculation. After Z-score standardization, the ICC between duration and sRPE reached 0.614 (substantial agreement), although standardization does not imply direct interchangeability. Among the 12 single indicators, EWMA uncoupled ACWR had the highest observed AUC (0.946). The EWMA four-indicator combination (coupled ACWR + uncoupled ACWR + Diff + SD) had the highest observed internal AUC (0.958). Uncoupled ACWR and Diff indicators showed numerically higher AUCs than their coupled and SD counterparts, respectively; these differences were not formally tested statistically and confidence intervals overlapped. Conclusions: In this small, single-team, single-event sample of female adolescent 200 m individual medley swimmers, sRPE showed moderate-to-substantial agreement with traditional workload indicators. The EWMA four-indicator model showed the highest observed internal discriminative ability for same-week injury classification (AUC = 0.958), with EWMA uncoupled ACWR the highest-ranked single indicator (AUC = 0.946). These findings are preliminary because of the small number of injury events, class imbalance, athlete-level clustering, binary injury labels, absence of simpler-model comparison, and lack of lagged or external validation. The exploratory two-tier warning thresholds (yellow > 0.5, red > 0.7) are conceptual, non-optimized boundaries and require prospective derivation and validation in independent cohorts before any operational use.