Quantitative comparison of explainable AI methods for interpreting deep learning-based classification of 3D gait kinematics.

Gait disorders can be caused by various reasons including cerebral palsy and neuromuscular diseases. 3D clinical gait analysis (3DGA) serves as a valuable clinical tool to assess gait abnormalities. Our previous research introduced a diagnostic tool that combines deep learning (DL) with 3DGA to evaluate childhood gait disorders. It achieved a promising diagnostic accuracy ranging from 0.77 to 0.99 across different pathologies. However, the lack of transparency limits their adoption. This research seeks to unveil the critical features that drive these models' diagnoses, improving interpretability and building trust in their decision-making process. Four different explaining artificial intelligence (XAI) methods were applied: LIME, DeepLift, Integrated Gradients, and sequential feature selection. These methods were used on various network architectures applied to three separate datasets involving different gait disorders. The results show that the features highlighted by XAI methods are relevant and reliable for diagnostic purposes. Moreover, quantitative analysis indicated that Integrated Gradients is the most appropriate XAI method in this case. Further experiments demonstrate that using parts of the critical features can achieve better accuracy than using all of the features. In conclusion, this research identified the diagnostic basis of DL models through XAI methods, enhanced diagnostic accuracy by focusing on critical features, and improved clinicians' understanding and trust in the DL diagnostic tool.