(4) UTILITY OF MACHINE LEARNING FOR IDENTIFICATION OF ATHLETE AVAILABILITY PREDICTORS IN NCAA ELITE FEMALE ATHLETES

Purpose: Developing comprehensive modeling to identify athletes at risk of low athlete availability (AA) before injury would be of value to sport practitioners and could also be used to refine testing batteries and hone meaningful predictors to prioritize data streams. This study used machine learning to identify influential predictors (training load [TL], recovery, wellness, body composition [BC], and force plate jumps [FP]) of AA in female athletes. Methods: Fifty-two NCAA Division I lacrosse (LAX; n=28, age: 20.4±1.6yrs) and soccer (SOC; n=24, age: 20.2±1.5yrs) athletes were tracked across a season for AA (% of unmodified, completed practices and games). TL included season means of max speed (MS; km/h), distance (D; m), high-speed running (HSR; m≥15 km/h), acute to chronic workload ratios (ACWR; 7/28 day exponentially weighted moving average) for D and HSR, and top MS of the season. Wearable recovery rings captured sleep duration (sec), efficiency (%), resting heart rate (RHR; bpm), heart rate variability (ms), and temperature deviation from the mean (Δ°C). Daily wellness surveys were used to capture perceived ratings of stress, soreness, mood, fatigue, and sleep quality (-3 to +3 Likert scale), and hormonal profile. Body composition change scores (Δ; post-preseason; lean mass, fat-free mass index, bone mineral density) via total body dual energy x-ray absorptiometry were included. FP metrics included reactive strength index modified (RSImod; ms-1), eccentric mean braking (N), jump height (cm), and peak power (W). Influential predictors were identified using elastic net regression. Goodness of fit was described with root mean squared error (RMSE). Results: The combined model displayed an average AA of 85.4±23.5%. RMSEs for the combined, LAX, and SOC models were 17.8%, 8.9%, and 17.8%, respectively. Influential coefficients (β) for each model are found in Table 1. Significant predictors of TL, recovery, and wellness variables were selected in all models, with differing impact between teams. Separate team analyses demonstrated differing influential predictors; most notable contrasts included no FP variables influenced LAX AA, and no BC variables selected for SOC AA. Conclusions: Differences in RMSEs and predictors between teams suggest single-team models as a stronger approach to understanding key outcomes for AA. Practical Applications: For LAX, the positive FFMI and LM β suggest gaining FFM across the season promotes AA, possibly related to higher relative loads (D ACWR). The negative β of temperature and RHR indicate higher AA is associated with greater sleep recovery. Further, LAX athletes with faster MS and worse sleep quality are at a higher risk of injury. Avoiding acute reductions in TL and implementation of high speed exposure in training predicted higher AA for SOC, possibly affecting higher perceived soreness. Finally, higher RSImod may represent improved neuromuscular efficiency, an important aspect in resilience to injury for SOC. Acknowledgements: None