Data augmented Autoformer for runner injury prediction

Injury prevention remains an immense concern in professional sports, as the complex and various influencing factors hint the precise prediction of runner’s injury risk. Although the advancement of deep learning methodology provides a revolutionary impact on athletic data analysis, injury risk assessment and treatment design, imbalanced injury records and ambiguous interaction patterns limit the generalization and trustworthiness of existing solutions. This study proposes a novel model, VAE-Former, for accurate and robust injury forecasting with extremely imbalanced runner dataset. The proposed model introduces a variational autoencoder (VAE) model to extract latent features and generate synthetic injury record for efficient pattern recognition and a Transformer based model for multi-variant time series prediction. This data augmentation mechanism is beneficial for performance of predicting model by enriched features from the minority. Moreover, proposed model inherits the prominent capacity of Autoformer structure to formulate the hidden relationships from multi-variable time series. The series attention mechanism and decomposition module are expert in temporal dependence and cross correlation analysis for a clear link between contributors and injury risks. Experiments on a public runner injury dataset validate the remarkable performance of VAE-Former. Results of principal component analysis reveals the augmented data similar with original records. The involvement of latent features significantly improve forecasting accuracy of Autoformer. Compared with other solutions, VAE-Former achieves 79.76% accuracy in runner injury prediction and the involvement of data augmentation accomplishes about 19.2% accuracy increases. Our achievements declare a new perspective on athlete injury prediction with limited data conditions. The proposed model demonstrates promising performance for intelligent athlete management and robust risk assessment.