Ultramarathon Result and Injury Prediction using PyTorch

In this paper, it is explored how the data about trainings, competitions and weather can be leveraged for result and injury prediction in competitive running disciplines. As outcome, two prediction models based on multilayer neural networks are implemented using PyTorch framework for Python programming language and evaluated on dataset containing realistic data of ultramarathon runner. The first model treats problem as regression to predict the number of kilometers run for given ultramarathon duration, while another one determines whether injury will occur for given running distance or no using classification approach. According to our results, both models show satisfiable performance (up to 2% relative error for regression, 70% correct for classification), while the first one performs better, which can be explained by lack of enough injury records in the second case. Moreover, a companion mobile app developed using AppSheet and Google Apps Script for automated dataset construction is introduced.