714 FO46 – Predicting injury using strava data of runners training for the New York City Marathon

Background Runners frequently experience injuries, especially in the context of training for a marathon. Many runners collect training data through smartphone applications and GPS watches, but this data has not been used to predict injury. Objective To use training data collected through Strava to create and evaluate a predictive model for injury of runners during marathon training. Design Prospective observational trial. Setting 16 weeks leading up to the 2022 New York City Marathon. Participants 643 adult marathon registrants participated in the study and provided Strava data, mean age 43 (SD 12) and 343 (53%) female. Interventions A baseline survey was administered 16 weeks before the marathon, collecting demographics, running/marathon experience, and injury history. During the 16 weeks before the marathon, health surveillance surveys were collected weekly and training runs were logged using Strava. Main Outcome Measurements First occurrence of running injury that resulted in modification of training. A machine learning algorithm allowing for non-linear relationships (generalized additive model) was trained on a random 80% of runners and tested on the remaining 20%. Predictors of injury each week included baseline information, prior week fatigue and happiness, and prior week training information. Results 326 (51%) reported an injury during training. Figures 1A and 1B show model performance on the training set (AUROC=0.79) and holdout test set (AUROC=0.71). Figure 2 shows the top predictors of injury, including history of running injury, running distance 7:28-day acute:chronic workload ratio (ACWR), and fatigue. Figure 3 shows the non-linear relationship the model learned between ACWR and injury risk (i.e., a step-up in injury risk at ACWR=1.35). Conclusions A machine learning model predicted the likelihood of injury among runners training for a marathon with fair accuracy. Changes in training load and fatigue are modifiable risk factors that could inform future injury prevention programs.