Influence of Number of Subjects and Number of Trials on Biomechanical Variable Estimation via Deep-Learning Models and Wearable IMUs During Drop Landings

Abstract

Data diversity and quantity are crucial for training deep-learning models. However, the impact of dataset diversity and size on biomechanical variable estimation models has not been explicitly investigated during drop landings. This work investigates the impact of the number of subjects and the number of trials per subject on the performance of wearable inertial measurement unit (IMU)-driven deep-learning models for knee moment and ground reaction force estimation during drop-landing tasks. An investigation dataset with 16 subjects and 25 trials per subject was collected in a biomechanical laboratory. The impact of subject and trial quantification was explored under different model complexity and types, as well as data augmentation methods using the investigation dataset. The deep-learning models were implemented by a feature extractor and an estimator realized by several fully connected (FC) layers. The feature extractor was independently evaluated with FC neural networks, convolutional neural network (CNN), long short-term memory (LSTM) model, and transformer model. Three transformation-based data augmentation methods were proposed and compared with the measured dataset (MD). The results showed that the minimum required number of subjects and trials for the models to achieve an estimation performance of 0.85 of R-squared, 0.4 body weight $\times $ body height of root mean square error (RMSE), and 0.1 of relative RMSE (rRMSE) is five subjects and five trials. Intriguingly, adding more subjects to the dataset improved the estimation performance while adding more trials did not. In addition, the proposed data augmentation can alleviate the data scarcity issue when the number of trials is small.