901 BO51 – Cross-sectional study of machine learning and inertial measurement unit performance to classify upper limb movements in an elite women’s water polo practice session

Background Water polo shoulder injuries account for 25% of lost training days (Hams et al, 2018). Increased shooting volumes in practice sessions are associated with these injuries (Wheeler et al, 2013). External load monitoring informs practice session planning whilst minimising injury risk – however no method is developed for practice session use. Inertial measurement units (IMUs) and machine learning techniques can classify water polo upper limb movements during standardised baseline testing but it is unknown whether this method can be applied to women's practice sessions (King at al, 2021). Objective To investigate machine learning model performance using uniaxial and peak resultant acceleration (RAccPEAK) and angular velocity (AngVelPEAK) during one practice session – high intensity throwing (HIT), low intensity throwing (LIT), blocking with/without ball contact (BWBC/BWOBC) and swimming (SWIM) compared to coded video analysis Design Cross-sectional Setting Institute pool Participants Ten elite women's athletes – 21.2 (4.8) years, 8.7 (SD 4.6) training years Interventions Two Blue TridentTM IMUs were affixed to the wrist and upper back. Sensors collected kinematic variables during one two-hour practice session to maximise movement repetition and ecological validity. Session was digitally videoed and coded for movements. Data pre-processing, feature and model development were performed in R-Studio with data partitioned into training and test sets using a random forest algorithm. Model movement classification accuracy was compared to coded video. Main Outcome Measurements Movement-specific RAccPEAK and AngVelPEAK and model accuracy Results Movements demonstrated differing RAccPEAK and AngVelPEAK values which were then used in model development. Model classified 1570 movements with 128 false positives (HIT – 62, LIT – 426, BWBC – 1, BWOBC – 24, SWIM – 930) with 92% accuracy. Conclusion This method demonstrates potential as a practice load monitoring tool. Longitudinal external load monitoring during practice sessions could inform performance progression, session planning and injury risk minimisation.