Using DeepLabCut for tracking body landmarks in videos of children with dyskinetic cerebral palsy: a working methodology

Markerless motion tracking is a promising technique to capture human movements and postures. It could be a clinically feasible tool to objectively assess movement disorders within severe dyskinetic cerebral palsy (CP). Here, we aim to evaluate tracking accuracy on clinically recorded video data. Method 94 video recordings of 33 participants (dyskinetic CP, 8-23 years; GMFCS IV-V, i.e. non-ambulatory) from a previous clinical trial were used. Twenty-second clips were cut during lying down as this is a postion for this group of children and young adults allows to freely move. Video image resolution was 0.4 cm per pixel. Tracking was performed in DeepLabCut. We evaluated a model that was pre-trained on a human healthy adult data set with an increasing number of manually labeled frames (0, 1, 2, 6, 10, 15 and 20 frames per video). To assess generalizability, we used 80% of videos for the model development and evaluated the generalizability of the model using the remaining 20%. For evaluation the mean absolute error (MAE) between DeepLabCut’s prediction of the position of body points and manual labels was calculated. Results Using just the pre-trained adult human model yielded a MAE of 121 pixels. An MAE of 4.5 pixels (about 1.5 cm) could be achieved by adding 15-20 manual labels. When applied to unseen video clips (i.e. generalization set), the MAE was 33 pixels with a dedicated model trained on 20 frames per videos. Conclusion Accuracy of tracking with a standard pre-trained model is insufficiently to automatically assess movement disorders in dyskinetic CP. However, manually adding labels improves the model performance substantially. In addition, the methodology proposed within our study is applicable to check the accuracy of DeepLabCut application within other clinical data set.