Applying deep learning to IMU data to classify lameness location in horses

Lameness assessment in horses is challenging and could be improved with gait quantification systems and deep learning. This work evaluated inertial measurement unit (IMU) stride data classification as Sound, Front or Hind lame using convolutional neural network (CNN) and Fourier analysis threshold-based (FA) classifiers. Retrospective data from different unilateral lameness induction studies were used (Shoe, LPS and patellar ligament models). Twenty horses were trotted in a straight line with seven IMU sensors (200 Hz). Vertical displacements of the head (H), withers (W) and sacrum (S) were segmented into strides, then labelled as Sound (baseline), Front or Hind (successful front/hindlimb induction; median absolute Head-Difference-Min: 44 mm and Sacrum-Difference-Min/Max: 19/19 mm respectively). Horses were split into training, cross-validation, and testing sets (14-2-4) for CNN. The same training and test horses were used to define and evaluate the FA thresholds. CNN inputs were H-W-S, H-S or W-S strides, while FA used H-S or W-S. Performances were evaluated with mean F1-score per class (Sound, Front, Hind), over ten different training-(validation)-testing sets. CNN presented higher F1-scores for Sound classification than FA, where CNNH-W-S and CNNH-S performed the best (F1-score: [69,66]%). For Front classification, H-S outperformed W-S for both CNN and FA (F1-score: CNN:[74,49]% vs FA:[74,24]%). FA was not able to classify Hind strides (F1-score<50%), while CNNs had F1-scores>70%. This study shows that using upper-body displacements and CNN, it is possible to classify sound, front and hind strides. Our comparison approach can also aid in understanding which IMU locations are crucial for lameness detection and classification.