Comparative analysis of biomechanical patterns in sprinting: A machine learning approach to optimize running performance in track athletes

Abstract

Listen

Translate article

Athletes’ success in field and track competitions has been reported to be determined by their sprinting skills. Therefore, it is crucial to understand what biomechanical and physiological factors contribute to the most effective sprinting attributes. The scientific research on sprint evaluation has predominantly dealt with discrete metrics simultaneously, avoiding the interplay between multiple factors as the sprint progresses. Incorporating all the factors that could potentially influence the impact of excellent sprint ability is the primary objective of the present study. This research investigates the biomechanics of sprinting using a Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) hybrid approach, focusing mainly on factors like stride length, ground reaction forces, joint angles, and muscle activation patterns. The hybrid Machine Learning (ML) model accurately identifies between the two groups, and the results indicate that sprinters performing at the national level have more extended movements, higher reaction time to ground forces, and improved joint angles. The research project set up a 20-meter track for the race, and 30 participants, divided 50-50 between two distinct groups that included comparable college-level and national-level performers, participated. With a 92.4% accuracy, 90.2% precision, and 90.9% F1 score, the hybrid approach performed better than standard models in predicting optimum sprinting patterns. The higher efficiency is caused by phase-specific changes that the model unattended, such as enhanced knee angles and joint accelerated motion in the swing phase. In comparison, the SVM model, though respectable, lags behind with an accuracy of 85.7% and a lower precision and recall (82.4% and 80.9%, respectively). The RF model performed better than SVM with an accuracy of 88.1% and a balanced F1-score of 86.8% but still fell short of the CNN-LSTM hybrid. The standalone LSTM model performed relatively well, with an accuracy of 89.3% and an F1 score of 88.1%, showing its capability but still not matching the hybrid model’s performance.