Examining Machine Learning Imputation Methods On Athlete Monitoring Assessments In Female Collegiate Volleyball

Advances in the complexity of athlete monitoring methods has required more intricate analysis design to appropriately interpret data insights. Missing data is a common challenge for practitioners when analyzing data, and machine learning (ML) imputation may be useful tool to create more robust analyses. Use of ML may aid to identify trends otherwise missed with traditional measurement considerations. PURPOSE: To examine the performance of ML data imputations on different sport related variables in collegiate female volleyball players during competition and preparation seasons. METHODS: Participants were twenty-two college-aged female volleyball players (Height = 176.6 ± 6.77 cm, Weight = 71.9 ± 5.2 kg) from different cohorts at the same institution over 5 years. During the preparatory (PREP) and competitive (COMP) seasons, the following variables were assessed: vertical jump (VJ), vertical jump with approach (VJA), 20-meter sprint (SPD), and measures of arm [bench press (BP), lat. pull (LP)] and leg strength [squat (SQ), leg curl (LC)]. Imputation methods were performed by randomly removing 30% of original data across each variable and assessed against the complete data (CON). Impute methods consisted of multiple imputation using chain sequence equations (MICE), k-nearest neighbors (kNN), and random forest methods (RF). Imputation (IMP) performance of respective method was assessed. A two factor ANOVA (IMP/season) was conducted to compare effects between IMP and CON, and IMP between seasons effects on monitored variables. RESULTS: Significant main effects existed for season on SQ [F (1, 1047) = 3.92, p = .048], LP [F (1, 1047) = 5.14, p = .023], VJA [F (1, 1047) = 8.32, p = .004], and SPD [F (1, 1047) = 25.00, p < .001]; with SQ, LP, VJA, and SPD performances being better in competitive than preparatory seasons. There were no significant main effects between IMP and CON (p > .05), and no interaction effect between IMP and season (p > .05) on any monitored variables. CONCLUSIONS: These results suggest that ML imputation methods could be utilized in athlete monitoring applications when up to 30% of data is missing in related variables monitored. Use of ML techniques could be an effective solution for data analyses due to varying conditions affecting sampling for performance practitioners’ decision-making on periodization.