Vertical loading rate prediction using artificial neural network

![CDATA[Rational & Objective: Vertical loading rate (VLR) is highly associated with running injuries. However, clinical tools that can easily assess VLR are still lacking. In the present study, we sought to develop a model relating peak vertical tibial acceleration, body weight and running speed with VLR using artificial neural network (ANN). Methods: Twenty regular runners were recruited from local running clubs. Two inertial measuring units (IMU) were firmly affixed onto the medial distal tibia bilaterally. All the participants were asked to run on an instrumented treadmill at 7 different speeds ranging from 7 km/h to 15 km/h (1 km/h increment). We collected synchronized IMU and kinetics data for the last 30 seconds from each 5-minute running trial. A two-layer feed-forward ANN with 10 sigmoid hidden neurons and linear output neurons was trained using peak vertical tibial acceleration, body weight and running speed as input and VLR as target. Bayesian regularization algorithm was used for back-propagation. 70% of the dataset was used as the training set while the rest of the data was divided into validation set and testing set. Results: The final output of the trained ANN has a linear fit with VLR of R squared of 0.9090 for the training dataset, while the model achieved similar fit with R squared of 0.9196 and 0.9110 for validation and testing dataset respectively, indicating that the model is generalizable. Conclusions: Our findings indicated that by obtaining data from wearable sensors (e.g. IMU), running speed and body weight, an accurate model without over-fitting can already be constructed using deep learning approach.]]