Identifying Key Training Load and Intensity Indicators in Ice Hockey Using Unsupervised Machine Learning

To identify key training load (TL) and intensity indicators in ice hockey, practice, and game data were collected using a wearable 200-Hz accelerometer and heart rate (HR) recording throughout a four-week (29 days) competitive period (23 practice sessions and 8 competitive games in 17 elite Danish players (n = 427 observations). Within-individual correlations among accelerometer- (total accelerations [Acc_tot], accelerations >2 m·s^-2 [Acc2], total accelerations [Dec_tot], decelerations <- 2 m·s^-2 [Dec2]), among HR-derived (time >85% maximum HR [t85%HR_max], Edwards' TL and modified training impulse) TL indicators, and between acceleration- and HR-derived TL parameters were large to almost perfect (r = 0.69-0.99). No significant correlations were observed between accelerometer- and HR-derived intensity indicators. Three between- and two within-components were found. The K-means++ cluster analysis revealed five and four clusters for between- and within-loadings, respectively. The least Euclidean distance from their centroid for each cluster was reported by session-duration, Acc_tot, Dec2, TRIMP_MOD, %t85HR_max for between-loadings, whereas session-duration, Acc2, t85HR_max and Dec2/min for within-loadings. Specific TL or intensity variables might be relevant to identify similar between-subject groups (e.g. individual player, playing positions), or temporal patterns (e.g. changes in TL or intensity over time). Our study provides insights about the redundancy associated with the use of multiple TL and intensity variables in ice hockey.