Forecasting Runner Injuries from Wearable Data using Recurrent Neural Networks
Date:
Wearable devices provide rich longitudinal data on training load and wellness, making data-driven injury prevention an emerging priority in sports science. In this project, we forecasted running-related injuries using multi-day sequences of workload and perceived-recovery metrics collected from wearable sensors. We processed ~40,000 daily windows of athlete training history, constructed multivariate time-series datasets, and trained a series of recurrent neural network models to identify temporal patterns preceding injury events. Our bidirectional gated recurrent unit model learns temporal signals associated with injury-prone periods and achieves stable predictive performance despite extreme class imbalance. Ongoing work includes evaluating more advanced sequence architectures and exploring data augmentation methods to improve minority-class representation. This work demonstrates the feasibility of leveraging wearable-derived time-series data for proactive injury forecasting, informing early-warning tools for athlete health and performance.