Vers des outils de diagnostic crédibles pour les troubles neurologiques de la marche : une fusion de la cinématique 3D, de l’apprentissage profond et de l’IA explicable

Gait abnormalities are common in children and can result from various pathologies. The analysis of gait patterns through 3D gait analysis provides clinically valuable information. In this study, we explored whetherdeep learning techniques could be combined with 3D gait data to diagnose gait disorders in children. We evaluated the diagnostic accuracy of several deep learning models using six datasets that included 3D gait data from children with five distinct pathologies and typically developing children. The models demonstrated good to excellent diagnostic accuracy, ranging from 0.77 to 0.99, across various diagnostic tasks. Our findings indicate that gait patterns contain pathology-specific information, suggesting their potential use in diagnostic applications. Additionally, we applied multiple explainable AI (XAI) methods to identify the key features driving model predictions, thus improving the interpretability and trustworthiness of the diagnostic models. The XAI results revealed that the highlighted features were both relevant and reliable for diagnostic purposes. Furthermore, our experiments demonstrated that focusing on the most informative features improved the diagnostic accuracy of the deep learning models. This study represents a first step toward betterunderstanding gait patterns in children. Future work will be needed to refine these models and further enhance their diagnostic capabilities.