Creative Commons source reading · Load, readiness and fatigue
Explainable temporal deep learning for athlete-independent classification of injury-labeled days in competitive distance runners: a methodological benchmark using seven-day training-load histories
Qiao T; Tian W
Originally published in Front Public Health · 2026-08-28
From the article
Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.
Abstract
AbstractIntroductionThis study evaluates whether seven-day training-load histories can support athlete-independent classification of injury-labeled days in competitive distance runners. The aim is methodological benchmarking rather than clinical injury-risk deployment.Research gapPrior sports-injury prediction studies often use repeated athlete-day observations without fully addressing athlete leakage, rare-event imbalance, probability calibration, and the risk that explainability methods may identify recording artifacts rather than physiological mechanisms. Data and method: The public dataset contained 42,766 athlete-day observations from 74 runners, including 583 injury-labeled observations (1.36%). Seven external-load variables and three internal-response variables were organized from Day −7 to Day −1. A compact dual-stream…
Cohort and descriptive results
The daily analysis included 42,766 athlete-day records from 74 runners. There were 583 injury-labeled observations and 42,183 non-injury observations, corresponding to 1.363% prevalence. The imbalance was therefore approximately 72 non-injury observations per injury-labeled observation. Follow-up differed markedly between athletes, and several athletes had few or no injury-labeled observation. At the athlete level, mean seven-day session count was 6.185 in injury-labeled categories and 5.499 in no-injury categories. Corresponding mean total distance was 54.893 vs. 45.990 km, moderate-intensity distance was 6.697 vs. 4.731 km, higher-intensity distance was 4.287 vs. 3.531 km, and sprint distance was 1.266 vs. 0.637 km. Mean perceived exertion, training success, and recovery were also descriptively higher in injury-labeled categories.…
Ablation and fusion results
Across five seeds, the ungated dual-stream model achieved the highest mean ROC-AUC among the deep variants (0.604 ± 0.011), followed by the full gated model (0.598 ± 0.007), internal-only model (0.575 ± 0.029), and external-only model (0.538 ± 0.021). The gated architecture therefore did not improve discrimination over simple concatenation, and H2 was rejected. The mean external-stream gate values remained close to 0.5 across held-out folds and seeds. A value near 0.5 indicates that the learned fusion behaved largely as near-uniform averaging rather than strong feature-wise selection of one stream over the other. The gate did not collapse completely to a single stream, but its near-balanced values and absence of incremental performance do not support a claim of effective selective fusion (Table 6).
Editorial reading guide
- How did the authors define training exposure and performance response?
- What alternative explanations, confounders or individual differences limit the conclusion?
