Prediction of continuous and discrete kinetic parameters in horses from inertial measurement units data using recurrent artificial neural networks.

The study uses artificial neural networks to predict Ground Reaction Force (GRFz) measurements in horses from Inertial Measurement Units (IMUs) data. The goal is to easier assess weight-bearing lameness in horses, which may enable better, less invasive, and more practical clinical use.

Methodology

• The researchers used a group of 24 clinically sound horses for their experiment. These horses were equipped with IMUs on their upper bodies and on each limb.
• The horses walked and trotted on a Ground Reaction Force (GRFz) treadmill – a device used to measure the force exerted by the horse on the treadmill surface.
• The data from these devices was time-synchronized for accuracy.
• The horses were split into groups of 16, 4, and 4 which were used for training, validation, and testing datasets respectively.
• These datasets were then fed into Long Short-Term Memory Recurrent Neural Networks (LSTM-RNNs). These are a type of artificial neural network known for their ability to make predictions based on time series data.
• The LSTM-RNNs trained on different input sets (All, Limbs, UB, Sacrum, or Withers) to predict the GRFz curves, or peak-GRFz values.

Results

• The study found that their models could predict the shape of the GRFz recently enough (below 0.40 N.kg for the Root Mean Square Error) at both the walking and trotting gaits.
• The best predictions of peak-GRFz values were achieved by utilizing data from all IMUs.
• When predicting both GRFz curves and peak-GRFz values, the models trained on all the data or just the upper body data consistently performed better than models trained on just the limbs’ data.
• This indicates the importance of kinematic data from the upper body in making accurate predictions of kinetic parameters.

Conclusion

• The successful performance of the LSTM-RNN models in this study suggests that this approach could be a viable method for more practical assessment of weight-bearing lameness in horses.
• However, the researchers caution that more data is required to confirm the effectiveness of this approach, as GRFz predictions can depend heavily on factors such as the speed of the horse, its gait, and the presence of any weight-bearing lameness.