Cross-site model development and validation of the Shriners Gait Index

Instrumented gait analysis is commonly used to aid clinicians in the development of treatment plans for children with cerebral palsy. The Shriners Gait Index has been developed as a deep learning autoencoder approach to present a holistic view of a person’s gait quality. This index is a single number which accounts for a substantial amount of variability of 114 clinically relevant gait parameters across subjects. The initial model was developed using data from a single site with model variants to account for kinetics and more severe gait impairments. With the availability of additional data from other sites, we are able to perform external validation of the model and its variants. To perform external validation, we applied the model parameters over the new data and compared the results. The external validation showed consistency across the multiple sites with the explained variance and reconstruction errors. The SGI accounted for approximately 92% of the variance across all scores for the base model, approximately 89% when the kinetics were included and approximately 88% when including more severely impaired individuals. These results indicate that the models will perform well on unseen data.We then included data from additional sites to study its effect on model performance. Multi-site models were tuned then trained with data from multiple sites. The multi-site models showed an average reduction of 6% of the unexplained variance over the initial models, showing that having more varied data for training allows the gait index to better represent a holistic account of gait impairments.